0bef826a98
Support llava clip vision model.
2025-03-06 00:24:43 -05:00
85ef295069
Make applying embeddings more efficient.
...
Adding new tokens no longer makes a whole copy of the embeddings weight
which can be massive on certain models.
2025-03-05 17:34:38 -05:00
5d84607bf3
Add type hint for FileLocator ( #6968 )
...
* Add type hint for FileLocator
* nit
2025-03-05 15:35:26 -05:00
c1909f350f
Better argument handling of front-end-root ( #7043 )
...
* Better argument handling of front-end-root
Improves handling of front-end-root launch argument. Several instances where users have set it and ComfyUI launches as normal and completely disregards the launch arg which doesn't make sense. Better to indicate to user that something is incorrect.
* Removed unused import
There was no real reason to use "Optional" typing in ther front-end-root argument.
2025-03-05 15:34:22 -05:00
52b3469606
[NodeDef] Explicitly add control_after_generate to seed/noise_seed ( #7059 )
...
* [NodeDef] Explicitly add control_after_generate to seed/noise_seed
* Update comfy/comfy_types/node_typing.py
Co-authored-by: filtered <176114999+webfiltered@users.noreply.github.com>
---------
Co-authored-by: filtered <176114999+webfiltered@users.noreply.github.com>
2025-03-05 15:33:23 -05:00
369b079ff6
Fix lowvram issue with ltxv vae.
2025-03-05 05:26:08 -05:00
9c9a7f012a
Adjust ltxv memory factor.
2025-03-05 05:16:05 -05:00
93fedd92fe
Support LTXV 0.9.5.
...
Credits: Lightricks team.
2025-03-05 00:13:49 -05:00
65042f7d39
Make it easier to set a custom template for hunyuan video.
2025-03-04 09:26:05 -05:00
7c7c70c400
Refactor skyreels i2v code.
2025-03-04 00:15:45 -05:00
f86c724ef2
Temporal area composition.
...
New ConditioningSetAreaPercentageVideo node.
2025-03-03 06:50:31 -05:00
9af6320ec9
Make 2d area composition nodes work on video models.
2025-03-02 08:19:16 -05:00
4dc6709307
Rename argument in last commit and document the options.
2025-03-01 02:43:49 -05:00
4d55f16ae8
Use enum list for --fast options ( #7024 )
2025-03-01 02:37:35 -05:00
cf0b549d48
--fast now takes a number as argument to indicate how fast you want it.
...
The idea is that you can indicate how much quality vs speed you want.
At the moment:
--fast 2 enables fp16 accumulation if your pytorch supports it.
--fast 5 enables fp8 matrix mult on fp8 models and the optimization above.
--fast without a number enables all optimizations.
2025-02-28 02:48:20 -05:00
eb4543474b
Use fp16 for intermediate for fp8 weights with --fast if supported.
2025-02-28 02:17:50 -05:00
1804397952
Use fp16 if checkpoint weights are fp16 and the model supports it.
2025-02-27 16:39:57 -05:00
f4dac8ab6f
Wan code small cleanup.
2025-02-27 07:22:42 -05:00
89253e9fe5
Support Cambricon MLU ( #6964 )
...
Co-authored-by: huzhan <huzhan@cambricon.com>
2025-02-26 20:45:13 -05:00
3ea3bc8546
Fix wan issues when prompt length is long.
2025-02-26 20:34:02 -05:00
0270a0b41c
Reduce artifacts on Wan by doing the patch embedding in fp32.
2025-02-26 16:59:26 -05:00
c37f15f98e
Add fast preview support for Wan models.
2025-02-26 08:56:23 -05:00
4bca7367f3
Don't try to use clip_fea on t2v model.
2025-02-26 08:38:09 -05:00
b6fefe686b
Better wan memory estimation.
2025-02-26 07:51:22 -05:00
fa62287f1f
More code reuse in wan.
...
Fix bug when changing the compute dtype on wan.
2025-02-26 05:22:29 -05:00
0844998db3
Slightly better wan i2v mask implementation.
2025-02-26 03:49:50 -05:00
4ced06b879
WIP support for Wan I2V model.
2025-02-26 01:49:43 -05:00
cb06e9669b
Wan seems to work with fp16.
2025-02-25 21:37:12 -05:00
9a66bb972d
Make wan work with all latent resolutions.
...
Cleanup some code.
2025-02-25 19:56:04 -05:00
ea0f939df3
Fix issue with wan and other attention implementations.
2025-02-25 19:13:39 -05:00
f37551c1d2
Change wan rope implementation to the flux one.
...
Should be more compatible.
2025-02-25 19:11:14 -05:00
63023011b9
WIP support for Wan t2v model.
2025-02-25 17:20:35 -05:00
f40076096e
Cleanup some lumina te code.
2025-02-25 04:10:26 -05:00
96d891cb94
Speedup on some models by not upcasting bfloat16 to float32 on mac.
2025-02-24 05:41:32 -05:00
ace899e71a
Prioritize fp16 compute when using allow_fp16_accumulation
2025-02-23 04:45:54 -05:00
aff16532d4
Remove some useless code.
2025-02-22 04:45:14 -05:00
072db3bea6
Assume the mac black image bug won't be fixed before v16.
2025-02-21 20:24:07 -05:00
a6deca6d9a
Latest mac still has the black image bug.
2025-02-21 20:14:30 -05:00
41c30e92e7
Let all model memory be offloaded on nvidia.
2025-02-21 06:32:21 -05:00
12da6ef581
Apparently directml supports fp16.
2025-02-20 09:30:24 -05:00
c5be423d6b
Fix link pointing to non-exisiting docs ( #6891 )
...
* Fix link pointing to non-exisiting docs
The current link is pointing to a path that does not exist any longer.
I changed it to point to the currect correct path for custom nodes datatypes.
* Update node_typing.py
2025-02-20 07:07:07 -05:00
5715be2ca9
Fix Hunyuan unet config detection for some models. ( #6877 )
...
The change to support 32 channel hunyuan models is missing the `key_prefix` on the key.
This addresses a complain in the comments of acc152b674
2025-02-19 07:14:45 -05:00
afc85cdeb6
Add Load Image Output node ( #6790 )
...
* add LoadImageOutput node
* add route for input/output/temp files
* update node_typing.py
* use literal type for image_folder field
* mark node as beta
2025-02-18 17:53:01 -05:00
acc152b674
Support loading and using SkyReels-V1-Hunyuan-I2V ( #6862 )
...
* Support SkyReels-V1-Hunyuan-I2V
* VAE scaling
* Fix T2V
oops
* Proper latent scaling
2025-02-18 17:06:54 -05:00
b07258cef2
Fix typo.
...
Let me know if this slows things down on 2000 series and below.
2025-02-18 07:28:33 -05:00
31e54b7052
Improve AMD arch detection.
2025-02-17 04:53:40 -05:00
8c0bae50c3
bf16 manual cast works on old AMD.
2025-02-17 04:42:40 -05:00
530412cb9d
Refactor torch version checks to be more future proof.
2025-02-17 04:36:45 -05:00
e2919d38b4
Disable bf16 on AMD GPUs that don't support it.
2025-02-16 05:46:10 -05:00
1cd6cd6080
Disable pytorch attention in VAE for AMD.
2025-02-14 05:42:14 -05:00
d7b4bf21a2
Auto enable mem efficient attention on gfx1100 on pytorch nightly 2.7
...
I'm not not sure which arches are supported yet. If you see improvements in
memory usage while using --use-pytorch-cross-attention on your AMD GPU let
me know and I will add it to the list.
2025-02-14 04:18:14 -05:00
019c7029ea
Add a way to set a different compute dtype for the model at runtime.
...
Currently only works for diffusion models.
2025-02-13 20:34:03 -05:00
8773ccf74d
Better memory estimation for ROCm that support mem efficient attention.
...
There is no way to check if the card actually supports it so it assumes
that it does if you use --use-pytorch-cross-attention with yours.
2025-02-13 08:32:36 -05:00
1d5d6586f3
Fix ruff.
2025-02-12 06:49:16 -05:00
35740259de
mix_ascend_bf16_infer_err ( #6794 )
2025-02-12 06:48:11 -05:00
ab888e1e0b
Add add_weight_wrapper function to model patcher.
...
Functions can now easily be added to wrap/modify model weights.
2025-02-12 05:55:35 -05:00
d9f0fcdb0c
Cleanup.
2025-02-11 17:17:03 -05:00
b124256817
Fix for running via DirectML ( #6542 )
...
* Fix for running via DirectML
Fix DirectML empty image generation issue with Flux1. add CPU fallback for unsupported path. Verified the model works on AMD GPUs
* fix formating
* update casual mask calculation
2025-02-11 17:11:32 -05:00
af4b7c91be
Make --force-fp16 actually force the diffusion model to be fp16.
2025-02-11 08:33:09 -05:00
4027466c80
Make lumina model work with any latent resolution.
2025-02-10 00:24:20 -05:00
095d867147
Remove useless function.
2025-02-09 07:02:57 -05:00
caeb27c3a5
res_multistep: Fix cfgpp and add ancestral samplers ( #6731 )
2025-02-08 19:39:58 -05:00
3d06e1c555
Make error more clear to user.
2025-02-08 18:57:24 -05:00
43a74c0de1
Allow FP16 accumulation with `--fast` ( #6453 )
...
Currently only applies to PyTorch nightly releases. (>=20250208)
2025-02-08 17:00:56 -05:00
079eccc92a
Don't compress http response by default.
...
Remove argument to disable it.
Add new --enable-compress-response-body argument to enable it.
2025-02-07 03:29:21 -05:00
14880e6dba
Remove some useless code.
2025-02-06 05:00:37 -05:00
37cd448529
Set the shift for Lumina back to 6.
2025-02-05 14:49:52 -05:00
94f21f9301
Upcasting rope to fp32 seems to make no difference in this model.
2025-02-05 04:32:47 -05:00
60653004e5
Use regular numbers for rope in lumina model.
2025-02-05 04:17:25 -05:00
a57d635c5f
Fix lumina 2 batches.
2025-02-04 21:48:11 -05:00
8ac2dddeed
Lower the default shift of lumina to reduce artifacts.
2025-02-04 06:50:37 -05:00
3e880ac709
Fix on python 3.9
2025-02-04 04:20:56 -05:00
e5ea112a90
Support Lumina 2 model.
2025-02-04 04:16:30 -05:00
44e19a28d3
Use maximum negative value instead of -inf for masks in text encoders.
...
This is probably more correct.
2025-02-02 09:46:00 -05:00
0a0df5f136
better guide message for sageattention ( #6634 )
2025-02-02 09:26:47 -05:00
24d6871e47
add disable-compres-response-body cli args; add compress middleware; ( #6672 )
2025-02-02 09:24:55 -05:00
9e1d301129
Only use stable cascade lora format with cascade model.
2025-02-01 06:35:22 -05:00
8d8dc9a262
Allow batch of different sigmas when noise scaling.
2025-01-30 06:49:52 -05:00
222f48c0f2
Allow changing folder_paths.base_path via command line argument. ( #6600 )
...
* Reimpl. CLI arg directly inside folder_paths.
* Update tests to use CLI arg mocking.
* Revert last-minute refactor.
* Fix test state polution.
2025-01-29 08:06:28 -05:00
13fd4d6e45
More friendly error messages for corrupted safetensors files.
2025-01-28 09:41:09 -05:00
255edf2246
Lower minimum ratio of loaded weights on Nvidia.
2025-01-27 05:26:51 -05:00
67feb05299
Remove redundant code.
2025-01-25 19:04:53 -05:00
14ca5f5a10
Remove useless code.
2025-01-24 06:15:54 -05:00
96e2a45193
Remove useless code.
2025-01-23 05:56:23 -05:00
dfa2b6d129
Remove unused function lcm in conds.py ( #6572 )
2025-01-23 05:54:09 -05:00
d6bbe8c40f
Remove support for python 3.8.
2025-01-22 17:04:30 -05:00
e857dd48b8
Add gradient estimation sampler ( #6554 )
2025-01-22 05:29:40 -05:00
fb2ad645a3
Add FluxDisableGuidance node to disable using the guidance embed.
2025-01-20 14:50:24 -05:00
d8a7a32779
Cleanup old TODO.
2025-01-20 03:44:13 -05:00
ebf038d4fa
Use `torch.special.expm1` ( #6388 )
...
* Use `torch.special.expm1`
This function provides greater precision than `exp(x) - 1` for small values of `x`.
Found with TorchFix https://github.com/pytorch-labs/torchfix/
* Use non-alias
2025-01-19 04:54:32 -05:00
b1a02131c9
Remove comfy.samplers self-import ( #6506 )
2025-01-18 17:49:51 -05:00
507199d9a8
Uni pc sampler now works with audio and video models.
2025-01-18 05:27:58 -05:00
2f3ab40b62
Add warning when using old pytorch versions.
2025-01-17 18:47:27 -05:00
0aa2368e46
Fix some cosmos fp8 issues.
2025-01-16 17:45:37 -05:00
cca96a85ae
Fix cosmos VAE failing with videos longer than 121 frames.
2025-01-16 16:30:06 -05:00
31831e6ef1
Code refactor.
2025-01-16 07:23:54 -05:00
88ceb28e20
Tweak hunyuan memory usage factor.
2025-01-16 06:31:03 -05:00
23289a6a5c
Clean up some debug lines.
2025-01-16 04:24:39 -05:00
9d8b6c1f46
More accurate memory estimation for cosmos and hunyuan video.
2025-01-16 03:48:40 -05:00
6320d05696
Slightly lower hunyuan video memory usage.
2025-01-16 00:23:01 -05:00
25683b5b02
Lower cosmos diffusion model memory usage.
2025-01-15 23:46:42 -05:00
4758fb64b9
Lower cosmos VAE memory usage by a bit.
2025-01-15 22:57:52 -05:00
008761166f
Optimize first attention block in cosmos VAE.
2025-01-15 21:48:46 -05:00
cba58fff0b
Remove unsafe embedding load for very old pytorch.
2025-01-15 04:32:23 -05:00
2feb8d0b77
Force safe loading of files in torch format on pytorch 2.4+
...
If this breaks something for you make an issue.
2025-01-15 03:50:27 -05:00
c78a45685d
Rewrite res_multistep sampler and implement res_multistep_cfg_pp sampler. ( #6462 )
2025-01-14 18:20:06 -05:00
3aaabb12d4
Implement Cosmos Image/Video to World (Video) diffusion models.
...
Use CosmosImageToVideoLatent to set the input image/video.
2025-01-14 05:14:10 -05:00
1f1c7b7b56
Remove useless code.
2025-01-13 03:52:37 -05:00
90f349f93d
Add res_multistep sampler from the cosmos code.
...
This sampler should work with all models.
2025-01-12 03:10:07 -05:00
6c9bd11fa3
Hooks Part 2 - TransformerOptionsHook and AdditionalModelsHook ( #6377 )
...
* Add 'sigmas' to transformer_options so that downstream code can know about the full scope of current sampling run, fix Hook Keyframes' guarantee_steps=1 inconsistent behavior with sampling split across different Sampling nodes/sampling runs by referencing 'sigmas'
* Cleaned up hooks.py, refactored Hook.should_register and add_hook_patches to use target_dict instead of target so that more information can be provided about the current execution environment if needed
* Refactor WrapperHook into TransformerOptionsHook, as there is no need to separate out Wrappers/Callbacks/Patches into different hook types (all affect transformer_options)
* Refactored HookGroup to also store a dictionary of hooks separated by hook_type, modified necessary code to no longer need to manually separate out hooks by hook_type
* In inner_sample, change "sigmas" to "sampler_sigmas" in transformer_options to not conflict with the "sigmas" that will overwrite "sigmas" in _calc_cond_batch
* Refactored 'registered' to be HookGroup instead of a list of Hooks, made AddModelsHook operational and compliant with should_register result, moved TransformerOptionsHook handling out of ModelPatcher.register_all_hook_patches, support patches in TransformerOptionsHook properly by casting any patches/wrappers/hooks to proper device at sample time
* Made hook clone code sane, made clear ObjectPatchHook and SetInjectionsHook are not yet operational
* Fix performance of hooks when hooks are appended via Cond Pair Set Props nodes by properly caching between positive and negative conds, make hook_patches_backup behave as intended (in the case that something pre-registers WeightHooks on the ModelPatcher instead of registering it at sample time)
* Filter only registered hooks on self.conds in CFGGuider.sample
* Make hook_scope functional for TransformerOptionsHook
* removed 4 whitespace lines to satisfy Ruff,
* Add a get_injections function to ModelPatcher
* Made TransformerOptionsHook contribute to registered hooks properly, added some doc strings and removed a so-far unused variable
* Rename AddModelsHooks to AdditionalModelsHook, rename SetInjectionsHook to InjectionsHook (not yet implemented, but at least getting the naming figured out)
* Clean up a typehint
2025-01-11 12:20:23 -05:00
ee8a7ab69d
Fast latent preview for Cosmos.
2025-01-11 04:41:24 -05:00
2ff3104f70
WIP support for Nvidia Cosmos 7B and 14B text to world (video) models.
2025-01-10 09:14:16 -05:00
129d8908f7
Add argument to skip the output reshaping in the attention functions.
2025-01-10 06:27:37 -05:00
ff838657fa
Cleaner handling of attention mask in ltxv model code.
2025-01-09 07:12:03 -05:00
2307ff6746
Improve some logging messages.
2025-01-08 19:05:22 -05:00
d0f3752e33
Properly calculate inner dim for t5 model.
...
This is required to support some different types of t5 models.
2025-01-07 17:33:03 -05:00
4209edf48d
Make a few more samplers deterministic.
2025-01-07 02:12:32 -05:00
d055325783
Document get_attr and get_model_object ( #6357 )
...
* Document get_attr and get_model_object
* Update model_patcher.py
* Update model_patcher.py
* Update model_patcher.py
2025-01-06 20:12:22 -05:00
916d1e14a9
Make ancestral samplers more deterministic.
2025-01-06 03:04:32 -05:00
c496e53519
In inner_sample, change "sigmas" to "sampler_sigmas" in transformer_options to not conflict with the "sigmas" that will overwrite "sigmas" in _calc_cond_batch ( #6360 )
2025-01-06 01:36:47 -05:00
d45ebb63f6
Remove old unused function.
2025-01-04 07:20:54 -05:00
9e9c8a1c64
Clear cache as often on AMD as Nvidia.
...
I think the issue this was working around has been solved.
If you notice that this change slows things down or causes stutters on
your AMD GPU with ROCm on Linux please report it.
2025-01-02 08:44:16 -05:00
0f11d60afb
Fix temporal tiling for decoder, remove redundant tiles. ( #6306 )
...
This commit fixes the temporal tile size calculation, and removes
a redundant tile at the end of the range when its elements are
completely covered by the previous tile.
Co-authored-by: Andrew Kvochko <a.kvochko@lightricks.com>
2025-01-01 16:29:01 -05:00
79eea51a1d
Fix and enforce all ruff W rules.
2025-01-01 03:08:33 -05:00
c0338a46a4
Fix unknown sampler error handling in calculate_sigmas function ( #6280 )
...
Modernize calculate_sigmas function
2024-12-31 17:33:50 -05:00
1c99734e5a
Add missing model_options param ( #6296 )
2024-12-31 14:46:55 -05:00
67758f50f3
Fix custom node type-hinting examples ( #6281 )
...
* Fix import in comfy_types doc / sample
* Clarify docstring
2024-12-31 03:41:09 -05:00
b7572b2f87
Fix and enforce no trailing whitespace.
2024-12-31 03:16:37 -05:00
a90aafafc1
Add kl_optimal scheduler ( #6206 )
...
* Add kl_optimal scheduler
* Rename kl_optimal_schedule to kl_optimal_scheduler to be more consistent
2024-12-30 05:09:38 -05:00
d9b7cfac7e
Fix and enforce new lines at the end of files.
2024-12-30 04:14:59 -05:00
3507870535
Add 'sigmas' to transformer_options so that downstream code can know about the full scope of current sampling run, fix Hook Keyframes' guarantee_steps=1 inconsistent behavior with sampling split across different Sampling nodes/sampling runs by referencing 'sigmas' ( #6273 )
2024-12-30 03:42:49 -05:00
a618f768e0
Auto reshape 2d to 3d latent for single image generation on video model.
2024-12-29 02:26:49 -05:00
b504bd606d
Add ruff rule for empty line with trailing whitespace.
2024-12-28 05:23:08 -05:00
d170292594
Remove some trailing white space.
2024-12-27 18:02:30 -05:00
9cfd185676
Add option to log non-error output to stdout ( #6243 )
...
* nit
* Add option to log non-error output to stdout
- No change to default behaviour
- Adds CLI argument: --log-stdout
- With this arg present, any logging of a level below logging.ERROR will be sent to stdout instead of stderr
2024-12-27 14:40:05 -05:00
4b5bcd8ac4
Closer memory estimation for hunyuan dit model.
2024-12-27 07:37:00 -05:00
ceb50b2cbf
Closer memory estimation for pixart models.
2024-12-27 07:30:09 -05:00
160ca08138
Use python 3.9 in launch test instead of 3.8
...
Fix ruff check.
2024-12-26 20:05:54 -05:00
c4bfdba330
Support ascend npu ( #5436 )
...
* support ascend npu
Co-authored-by: YukMingLaw <lymmm2@163.com>
Co-authored-by: starmountain1997 <guozr1997@hotmail.com>
Co-authored-by: Ginray <ginray0215@gmail.com>
2024-12-26 19:36:50 -05:00
ee9547ba31
Improve temporal VAE Encode (Tiled) math.
2024-12-26 07:18:49 -05:00
19a64d6291
Cleanup some mac related code.
2024-12-25 05:32:51 -05:00
b486885e08
Disable bfloat16 on older mac.
2024-12-25 05:18:50 -05:00
0229228f3f
Clean up the VAE dtypes code.
2024-12-25 04:50:34 -05:00
99a1fb6027
Make fast fp8 take a bit less peak memory.
2024-12-24 18:05:19 -05:00
73e04987f7
Prevent black images in VAE Decode (Tiled) node.
...
Overlap should be minimum 1 with tiling 2 for tiled temporal VAE decoding.
2024-12-24 07:36:30 -05:00
5388df784a
Add temporal tiling to VAE Encode (Tiled) node.
2024-12-24 07:10:09 -05:00
bc6dac4327
Add temporal tiling to VAE Decode (Tiled) node.
...
You can now do tiled VAE decoding on the temporal direction for videos.
2024-12-23 20:03:37 -05:00
15564688ed
Add a try except block so if torch version is weird it won't crash.
2024-12-23 03:22:48 -05:00
c6b9c11ef6
Add oneAPI device selector for xpu and some other changes. ( #6112 )
...
* Add oneAPI device selector and some other minor changes.
* Fix device selector variable name.
* Flip minor version check sign.
* Undo changes to README.md.
2024-12-23 03:18:32 -05:00
e44d0ac7f7
Make --novram completely offload weights.
...
This flag is mainly used for testing the weight offloading, it shouldn't
actually be used in practice.
Remove useless import.
2024-12-23 01:51:08 -05:00
56bc64f351
Comment out some useless code.
2024-12-22 23:51:14 -05:00
f7d83b72e0
fixed a bug in ldm/pixart/blocks.py ( #6158 )
2024-12-22 23:44:20 -05:00
80f07952d2
Fix lowvram issue with ltxv vae.
2024-12-22 23:20:17 -05:00
57f330caf9
Relax minimum ratio of weights loaded in memory on nvidia.
...
This should make it possible to do higher res images/longer videos by
further offloading weights to CPU memory.
Please report an issue if this slows down things on your system.
2024-12-22 03:06:37 -05:00
da13b6b827
Get rid of meshgrid warning.
2024-12-20 18:02:12 -05:00
c86cd58573
Remove useless code.
2024-12-20 17:50:03 -05:00
b5fe39211a
Remove some useless code.
2024-12-20 17:43:50 -05:00
e946667216
Some fixes/cleanups to pixart code.
...
Commented out the masking related code because it is never used in this
implementation.
2024-12-20 17:10:52 -05:00
d7969cb070
Replace print with logging ( #6138 )
...
* Replace print with logging
* nit
* nit
* nit
* nit
* nit
* nit
2024-12-20 16:24:55 -05:00
bddb02660c
Add PixArt model support ( #6055 )
...
* PixArt initial version
* PixArt Diffusers convert logic
* pos_emb and interpolation logic
* Reduce duplicate code
* Formatting
* Use optimized attention
* Edit empty token logic
* Basic PixArt LoRA support
* Fix aspect ratio logic
* PixArtAlpha text encode with conds
* Use same detection key logic for PixArt diffusers
2024-12-20 15:25:00 -05:00
418eb7062d
Support new LTXV VAE.
2024-12-20 04:38:29 -05:00
cac68ca813
Fix some more video tiled encode issues.
...
The downscale_ratio formula for the temporal had issues with some frame
numbers.
2024-12-19 23:14:03 -05:00
52c1d933b2
Fix tiled hunyuan video VAE encode issue.
...
Some shapes like 1024x1024 with tile_size 256 and overlap 64 had issues.
2024-12-19 22:55:15 -05:00
2dda7c11a3
More proper fix for the memory issue.
2024-12-19 16:21:56 -05:00
3ad3248ad7
Fix lowvram bug when using a model multiple times in a row.
...
The memory system would load an extra 64MB each time until either the
model was completely in memory or OOM.
2024-12-19 16:04:56 -05:00
c441048a4f
Make VAE Encode tiled node work with video VAE.
2024-12-19 05:31:39 -05:00
9f4b181ab3
Add fast previews for hunyuan video.
2024-12-18 18:24:23 -05:00
cbbf077593
Small optimizations.
2024-12-18 18:23:28 -05:00
ff2ff02168
Support old diffusion-pipe hunyuan video loras.
2024-12-18 06:23:54 -05:00
4c5c4ddeda
Fix regression in VAE code on old pytorch versions.
2024-12-18 03:08:28 -05:00
37e5390f5f
Add: --use-sage-attention to enable SageAttention.
...
You need to have the library installed first.
2024-12-18 01:56:10 -05:00
a4f59bc65e
Pick attention implementation based on device in llama code.
2024-12-18 01:30:20 -05:00
ca457f7ba1
Properly tokenize the template for hunyuan video.
2024-12-17 16:22:02 -05:00
cd6f615038
Fix tiled vae not working with some shapes.
2024-12-17 16:22:02 -05:00
e4e1bff605
Support diffusion-pipe hunyuan video lora format.
2024-12-17 07:14:21 -05:00
d6656b0c0c
Support llama hunyuan video text encoder in scaled fp8 format.
2024-12-17 04:19:22 -05:00
f4cdedea62
Fix regression with ltxv VAE.
2024-12-17 02:17:31 -05:00
39b1fc4ccc
Adjust used dtypes for hunyuan video VAE and diffusion model.
2024-12-16 23:31:10 -05:00
bda1482a27
Basic Hunyuan Video model support.
2024-12-16 19:35:40 -05:00
19ee5d9d8b
Don't expand mask when not necessary.
...
Expanding seems to slow down inference.
2024-12-16 18:22:50 -05:00
61b50720d0
Add support for attention masking in Flux ( #5942 )
...
* fix attention OOM in xformers
* allow passing attention mask in flux attention
* allow an attn_mask in flux
* attn masks can be done using replace patches instead of a separate dict
* fix return types
* fix return order
* enumerate
* patch the right keys
* arg names
* fix a silly bug
* fix xformers masks
* replace match with if, elif, else
* mask with image_ref_size
* remove unused import
* remove unused import 2
* fix pytorch/xformers attention
This corrects a weird inconsistency with skip_reshape.
It also allows masks of various shapes to be passed, which will be
automtically expanded (in a memory-efficient way) to a size that is
compatible with xformers or pytorch sdpa respectively.
* fix mask shapes
2024-12-16 18:21:17 -05:00
e83063bf24
Support conv3d in PatchEmbed.
2024-12-14 05:46:04 -05:00
4e14032c02
Make pad_to_patch_size function work on multi dim.
2024-12-13 07:22:05 -05:00
563291ee51
Enforce all pyflake lint rules ( #6033 )
...
* Enforce F821 undefined-name
* Enforce all pyflake lint rules
2024-12-12 19:29:37 -05:00
2cddbf0821
Lint and fix undefined names (1/N) ( #6028 )
2024-12-12 18:55:26 -05:00
60749f345d
Lint and fix undefined names (3/N) ( #6030 )
2024-12-12 18:49:40 -05:00
d9d7f3c619
Lint all unused variables ( #5989 )
...
* Enable F841
* Autofix
* Remove all unused variable assignment
2024-12-12 17:59:16 -05:00
fd5dfb812c
Set initial load devices for te and model to mps device on mac.
2024-12-12 06:00:31 -05:00
7a7efe8424
Support loading some checkpoint files with nested dicts.
2024-12-11 08:04:54 -05:00
44db978531
Fix a few things in text enc code for models with no eos token.
2024-12-10 23:07:26 -05:00
1c8d11e48a
Support different types of tokenizers.
...
Support tokenizers without an eos token.
Pass full sentences to tokenizer for more efficient tokenizing.
2024-12-10 15:03:39 -05:00
23827ca312
Add `cond_scale` to `sampler_post_cfg_function` ( #5985 )
2024-12-09 20:13:18 -05:00
0fd4e6c778
Lint unused import ( #5973 )
...
* Lint unused import
* nit
* Remove unused imports
* revert fix_torch import
* nit
2024-12-09 15:24:39 -05:00
e2fafe0686
Make CLIP set last layer node work with t5 models.
2024-12-09 03:57:14 -05:00
fbf68c4e52
clamp input ( #5928 )
2024-12-07 14:00:31 -05:00
8af9a91e0c
A few improvements to #5937 .
2024-12-06 05:49:15 -05:00
005d2d3a13
ltxv: add noise to guidance image to ensure generated motion. ( #5937 )
2024-12-06 05:46:08 -05:00
1e21f4c14e
Make timestep ranges more usable on rectified flow models.
...
This breaks some old workflows but should make the nodes actually useful.
2024-12-05 16:40:58 -05:00
48272448ad
[Developer Experience] Add node typing ( #5676 )
...
* [Developer Experience] Add node typing
* Shim StrEnum
* nit
* nit
* nit
2024-12-04 15:01:00 -05:00
452179fe4f
Make ModelPatcher class clone function work with inheritance.
2024-12-03 13:57:57 -05:00
c1b92b719d
Some optimizations to euler a.
2024-12-03 06:11:52 -05:00
57e8bf6a9f
Fix case where a memory leak could cause crash.
...
Now the only symptom of code messing up and keeping references to a model
object when it should not will be endless prints in the log instead of the
next workflow crashing ComfyUI.
2024-12-02 19:49:49 -05:00
0ee322ec5f
ModelPatcher Overhaul and Hook Support ( #5583 )
...
* Added hook_patches to ModelPatcher for weights (model)
* Initial changes to calc_cond_batch to eventually support hook_patches
* Added current_patcher property to BaseModel
* Consolidated add_hook_patches_as_diffs into add_hook_patches func, fixed fp8 support for model-as-lora feature
* Added call to initialize_timesteps on hooks in process_conds func, and added call prepare current keyframe on hooks in calc_cond_batch
* Added default_conds support in calc_cond_batch func
* Added initial set of hook-related nodes, added code to register hooks for loras/model-as-loras, small renaming/refactoring
* Made CLIP work with hook patches
* Added initial hook scheduling nodes, small renaming/refactoring
* Fixed MaxSpeed and default conds implementations
* Added support for adding weight hooks that aren't registered on the ModelPatcher at sampling time
* Made Set Clip Hooks node work with hooks from Create Hook nodes, began work on better Create Hook Model As LoRA node
* Initial work on adding 'model_as_lora' lora type to calculate_weight
* Continued work on simpler Create Hook Model As LoRA node, started to implement ModelPatcher callbacks, attachments, and additional_models
* Fix incorrect ref to create_hook_patches_clone after moving function
* Added injections support to ModelPatcher + necessary bookkeeping, added additional_models support in ModelPatcher, conds, and hooks
* Added wrappers to ModelPatcher to facilitate standardized function wrapping
* Started scaffolding for other hook types, refactored get_hooks_from_cond to organize hooks by type
* Fix skip_until_exit logic bug breaking injection after first run of model
* Updated clone_has_same_weights function to account for new ModelPatcher properties, improved AutoPatcherEjector usage in partially_load
* Added WrapperExecutor for non-classbound functions, added calc_cond_batch wrappers
* Refactored callbacks+wrappers to allow storing lists by id
* Added forward_timestep_embed_patch type, added helper functions on ModelPatcher for emb_patch and forward_timestep_embed_patch, added helper functions for removing callbacks/wrappers/additional_models by key, added custom_should_register prop to hooks
* Added get_attachment func on ModelPatcher
* Implement basic MemoryCounter system for determing with cached weights due to hooks should be offloaded in hooks_backup
* Modified ControlNet/T2IAdapter get_control function to receive transformer_options as additional parameter, made the model_options stored in extra_args in inner_sample be a clone of the original model_options instead of same ref
* Added create_model_options_clone func, modified type annotations to use __future__ so that I can use the better type annotations
* Refactored WrapperExecutor code to remove need for WrapperClassExecutor (now gone), added sampler.sample wrapper (pending review, will likely keep but will see what hacks this could currently let me get rid of in ACN/ADE)
* Added Combine versions of Cond/Cond Pair Set Props nodes, renamed Pair Cond to Cond Pair, fixed default conds never applying hooks (due to hooks key typo)
* Renamed Create Hook Model As LoRA nodes to make the test node the main one (more changes pending)
* Added uuid to conds in CFGGuider and uuids to transformer_options to allow uniquely identifying conds in batches during sampling
* Fixed models not being unloaded properly due to current_patcher reference; the current ComfyUI model cleanup code requires that nothing else has a reference to the ModelPatcher instances
* Fixed default conds not respecting hook keyframes, made keyframes not reset cache when strength is unchanged, fixed Cond Set Default Combine throwing error, fixed model-as-lora throwing error during calculate_weight after a recent ComfyUI update, small refactoring/scaffolding changes for hooks
* Changed CreateHookModelAsLoraTest to be the new CreateHookModelAsLora, rename old ones as 'direct' and will be removed prior to merge
* Added initial support within CLIP Text Encode (Prompt) node for scheduling weight hook CLIP strength via clip_start_percent/clip_end_percent on conds, added schedule_clip toggle to Set CLIP Hooks node, small cleanup/fixes
* Fix range check in get_hooks_for_clip_schedule so that proper keyframes get assigned to corresponding ranges
* Optimized CLIP hook scheduling to treat same strength as same keyframe
* Less fragile memory management.
* Make encode_from_tokens_scheduled call cleaner, rollback change in model_patcher.py for hook_patches_backup dict
* Fix issue.
* Remove useless function.
* Prevent and detect some types of memory leaks.
* Run garbage collector when switching workflow if needed.
* Moved WrappersMP/CallbacksMP/WrapperExecutor to patcher_extension.py
* Refactored code to store wrappers and callbacks in transformer_options, added apply_model and diffusion_model.forward wrappers
* Fix issue.
* Refactored hooks in calc_cond_batch to be part of get_area_and_mult tuple, added extra_hooks to ControlBase to allow custom controlnets w/ hooks, small cleanup and renaming
* Fixed inconsistency of results when schedule_clip is set to False, small renaming/typo fixing, added initial support for ControlNet extra_hooks to work in tandem with normal cond hooks, initial work on calc_cond_batch merging all subdicts in returned transformer_options
* Modified callbacks and wrappers so that unregistered types can be used, allowing custom_nodes to have their own unique callbacks/wrappers if desired
* Updated different hook types to reflect actual progress of implementation, initial scaffolding for working WrapperHook functionality
* Fixed existing weight hook_patches (pre-registered) not working properly for CLIP
* Removed Register/Direct hook nodes since they were present only for testing, removed diff-related weight hook calculation as improved_memory removes unload_model_clones and using sample time registered hooks is less hacky
* Added clip scheduling support to all other native ComfyUI text encoding nodes (sdxl, flux, hunyuan, sd3)
* Made WrapperHook functional, added another wrapper/callback getter, added ON_DETACH callback to ModelPatcher
* Made opt_hooks append by default instead of replace, renamed comfy.hooks set functions to be more accurate
* Added apply_to_conds to Set CLIP Hooks, modified relevant code to allow text encoding to automatically apply hooks to output conds when apply_to_conds is set to True
* Fix cached_hook_patches not respecting target_device/memory_counter results
* Fixed issue with setting weights from hooks instead of copying them, added additional memory_counter check when caching hook patches
* Remove unnecessary torch.no_grad calls for hook patches
* Increased MemoryCounter minimum memory to leave free by *2 until a better way to get inference memory estimate of currently loaded models exists
* For encode_from_tokens_scheduled, allow start_percent and end_percent in add_dict to limit which scheduled conds get encoded for optimization purposes
* Removed a .to call on results of calculate_weight in patch_hook_weight_to_device that was screwing up the intermediate results for fp8 prior to being passed into stochastic_rounding call
* Made encode_from_tokens_scheduled work when no hooks are set on patcher
* Small cleanup of comments
* Turn off hook patch caching when only 1 hook present in sampling, replace some current_hook = None with calls to self.patch_hooks(None) instead to avoid a potential edge case
* On Cond/Cond Pair nodes, removed opt_ prefix from optional inputs
* Allow both FLOATS and FLOAT for floats_strength input
* Revert change, does not work
* Made patch_hook_weight_to_device respect set_func and convert_func
* Make discard_model_sampling True by default
* Add changes manually from 'master' so merge conflict resolution goes more smoothly
* Cleaned up text encode nodes with just a single clip.encode_from_tokens_scheduled call
* Make sure encode_from_tokens_scheduled will respect use_clip_schedule on clip
* Made nodes in nodes_hooks be marked as experimental (beta)
* Add get_nested_additional_models for cases where additional_models could have their own additional_models, and add robustness for circular additional_models references
* Made finalize_default_conds area math consistent with other sampling code
* Changed 'opt_hooks' input of Cond/Cond Pair Set Default Combine nodes to 'hooks'
* Remove a couple old TODO's and a no longer necessary workaround
2024-12-02 14:51:02 -05:00
79d5ceae6e
Improved memory management. ( #5450 )
...
* Less fragile memory management.
* Fix issue.
* Remove useless function.
* Prevent and detect some types of memory leaks.
* Run garbage collector when switching workflow if needed.
* Fix issue.
2024-12-02 14:39:34 -05:00
2d5b3e0078
Remove useless code.
2024-12-02 06:49:55 -05:00
8e4118c0de
make dpm_2_ancestral work with rectified flow.
2024-12-01 07:37:41 -05:00
26fb2c68e8
Add a way to disable cropping in the CLIPVisionEncode node.
2024-11-28 20:24:47 -05:00
bf2650a80e
Fast previews for ltxv.
2024-11-28 06:46:15 -05:00
b666539595
Remove print.
2024-11-27 20:28:39 -05:00
95d8713482
Missing parentheses.
2024-11-27 13:45:32 -05:00
497db6212f
Alternative fix for #5767
2024-11-26 17:53:04 -05:00
4c82741b54
Support official SD3.5 Controlnets.
2024-11-26 11:31:25 -05:00
15c39ea757
Support for the official mochi lora format.
2024-11-26 03:34:36 -05:00
b7143b74ce
Flux inpaint model does not work in fp16.
2024-11-26 01:33:01 -05:00
61196d8857
Add option to inference the diffusion model in fp32 and fp64.
2024-11-25 05:00:23 -05:00
b4526d3fc3
Skip layer guidance now works on hydit model.
2024-11-24 05:54:30 -05:00
ab885b33ba
Skip layer guidance node now works on LTX-Video.
2024-11-23 10:33:05 -05:00
839ed3368e
Some improvements to the lowvram unloading.
2024-11-22 20:59:15 -05:00
6e8cdcd3cb
Fix some tiled VAE decoding issues with LTX-Video.
2024-11-22 18:00:34 -05:00
e5c3f4b87f
LTXV lowvram fixes.
2024-11-22 17:17:11 -05:00
bc6be6c11e
Some fixes to the lowvram system.
2024-11-22 16:40:04 -05:00
5818f6cf51
Remove print.
2024-11-22 10:49:15 -05:00
5e16f1d24b
Support Lightricks LTX-Video model.
2024-11-22 08:46:39 -05:00
2fd9c1308a
Fix mask issue in some attention functions.
2024-11-22 02:10:09 -05:00
8f0009aad0
Support new flux model variants.
2024-11-21 08:38:23 -05:00
41444b5236
Add some new weight patching functionality.
...
Add a way to reshape lora weights.
Allow weight patches to all weight not just .weight and .bias
Add a way for a lora to set a weight to a specific value.
2024-11-21 07:19:17 -05:00
07f6eeaa13
Fix mask issue with attention_xformers.
2024-11-20 17:07:46 -05:00
22535d0589
Skip layer guidance now works on stable audio model.
2024-11-20 07:33:06 -05:00
b699a15062
Refactor inpaint/ip2p code.
2024-11-19 03:25:25 -05:00
d9f90965c8
Support block replace patches in auraflow.
2024-11-17 08:19:59 -05:00
41886af138
Add transformer options blocks replace patch to mochi.
2024-11-16 20:48:14 -05:00
3b9a6cf2b1
Fix issue with 3d masks.
2024-11-13 07:18:30 -05:00
8ebf2d8831
Add block replace transformer_options to flux.
2024-11-12 08:00:39 -05:00
eb476e6ea9
Allow 1D masks for 1D latents.
2024-11-11 14:44:52 -05:00
8b275ce5be
Support auto detecting some zsnr anime checkpoints.
2024-11-11 05:34:11 -05:00
2a18e98ccf
Refactor so that zsnr can be set in the sampling_settings.
2024-11-11 04:55:56 -05:00
bdeb1c171c
Fast previews for mochi.
2024-11-10 03:39:35 -05:00
8b90e50979
Properly handle and reshape masks when used on 3d latents.
2024-11-09 15:30:19 -05:00
2865f913f7
Free memory before doing tiled decode.
2024-11-07 04:01:24 -05:00
b49616f951
Make VAEDecodeTiled node work with video VAEs.
2024-11-07 03:47:12 -05:00
5e29e7a488
Remove scaled_fp8 key after reading it to silence warning.
2024-11-06 04:56:42 -05:00
8afb97cd3f
Fix unknown VAE being detected as the mochi VAE.
2024-11-05 03:43:27 -05:00
69694f40b3
fix dynamic shape export ( #5490 )
2024-11-04 14:59:28 -05:00
6c9dbde7de
Fix mochi all in one checkpoint t5xxl key names.
2024-11-03 01:40:42 -05:00
fabf449feb
Mochi VAE encoder.
2024-11-01 17:33:09 -04:00
1c8286a44b
Avoid SyntaxWarning in UniPC docstring ( #5442 )
2024-10-31 15:17:26 -04:00
1af4a47fd1
Bump up mac version for attention upcast bug workaround.
2024-10-31 15:15:31 -04:00
daa1565b93
Fix diffusers flux controlnet regression.
2024-10-30 13:11:34 -04:00
09fdb2b269
Support SD3.5 medium diffusers format weights and loras.
2024-10-30 04:24:00 -04:00
30c0c81351
Add a way to patch blocks in SD3.
2024-10-29 00:48:32 -04:00
13b0ff8a6f
Update SD3 code.
2024-10-28 21:58:52 -04:00
c320801187
Remove useless line.
2024-10-28 17:41:12 -04:00
669d9e4c67
Set default shift on mochi to 6.0
2024-10-27 22:21:04 -04:00
9ee0a6553a
float16 inference is a bit broken on mochi.
2024-10-27 04:56:40 -04:00
5cbb01bc2f
Basic Genmo Mochi video model support.
...
To use:
"Load CLIP" node with t5xxl + type mochi
"Load Diffusion Model" node with the mochi dit file.
"Load VAE" with the mochi vae file.
EmptyMochiLatentVideo node for the latent.
euler + linear_quadratic in the KSampler node.
2024-10-26 06:54:00 -04:00
c3ffbae067
Make LatentUpscale nodes work on 3d latents.
2024-10-26 01:50:51 -04:00
d605677b33
Make euler_ancestral work on flow models (credit: Ashen).
2024-10-25 19:53:44 -04:00
af8cf79a2d
support SimpleTuner lycoris lora for SD3 ( #5340 )
2024-10-24 01:18:32 -04:00
66b0961a46
Fix ControlLora issue with last commit.
2024-10-23 17:02:40 -04:00
754597c8a9
Clean up some controlnet code.
...
Remove self.device which was useless.
2024-10-23 14:19:05 -04:00
915fdb5745
Fix lowvram edge case.
2024-10-22 16:34:50 -04:00
5a8a48931a
remove attention abstraction ( #5324 )
2024-10-22 14:02:38 -04:00
8ce2a1052c
Optimizations to --fast and scaled fp8.
2024-10-22 02:12:28 -04:00
f82314fcfc
Fix duplicate sigmas on beta scheduler.
2024-10-21 20:19:45 -04:00
0075c6d096
Mixed precision diffusion models with scaled fp8.
...
This change allows supports for diffusion models where all the linears are
scaled fp8 while the other weights are the original precision.
2024-10-21 18:12:51 -04:00
83ca891118
Support scaled fp8 t5xxl model.
2024-10-20 22:27:00 -04:00
f9f9faface
Fixed model merging issue with scaled fp8.
2024-10-20 06:24:31 -04:00
471cd3eace
fp8 casting is fast on GPUs that support fp8 compute.
2024-10-20 00:54:47 -04:00
a68bbafddb
Support diffusion models with scaled fp8 weights.
2024-10-19 23:47:42 -04:00
73e3a9e676
Clamp output when rounding weight to prevent Nan.
2024-10-19 19:07:10 -04:00
67158994a4
Use the lowvram cast_to function for everything.
2024-10-17 17:25:56 -04:00
0bedfb26af
Revert "Fix Transformers FutureWarning ( #5140 )"
...
This reverts commit 95b7cf9bbe
2024-10-16 12:36:19 -04:00
f584758271
Cleanup some useless lines.
2024-10-14 21:02:39 -04:00
95b7cf9bbe
Fix Transformers FutureWarning ( #5140 )
...
* Update sd1_clip.py
Fix Transformers FutureWarning
* Update sd1_clip.py
Fix comment
2024-10-14 20:12:20 -04:00
3c60ecd7a8
Fix fp8 ops staying enabled.
2024-10-12 14:10:13 -04:00
7ae6626723
Remove useless argument.
2024-10-12 07:16:21 -04:00
6632365e16
model_options consistency between functions.
...
weight_dtype -> dtype
2024-10-11 20:51:19 -04:00
ad07796777
🐛 Add device to variable c ( #5210 )
2024-10-11 20:37:50 -04:00
1b80895285
Make clip loader nodes support loading sd3 t5xxl in lower precision.
...
Add attention mask support in the SD3 text encoder code.
2024-10-10 15:06:15 -04:00
5f9d5a244b
Hotfix for the div zero occurrence when memory_used_encode is 0 ( #5121 )
...
https://github.com/comfyanonymous/ComfyUI/issues/5069#issuecomment-2382656368
2024-10-09 23:34:34 -04:00
4b2f0d9413
Increase maximum macOS version to 15.0.1 when forcing upcast attention ( #5191 )
2024-10-09 22:21:41 -04:00
e38c94228b
Add a weight_dtype fp8_e4m3fn_fast to the Diffusion Model Loader node.
...
This is used to load weights in fp8 and use fp8 matrix multiplication.
2024-10-09 19:43:17 -04:00
203942c8b2
Fix flux doras with diffusers keys.
2024-10-08 19:03:40 -04:00
8dfa0cc552
Make SD3 fast previews a little better.
2024-10-07 09:19:59 -04:00
e5ecdfdd2d
Make fast previews for SDXL a little better by adding a bias.
2024-10-06 19:27:04 -04:00
7d29fbf74b
Slightly improve the fast previews for flux by adding a bias.
2024-10-06 17:55:46 -04:00
7d2467e830
Some minor cleanups.
2024-10-05 13:22:39 -04:00
6f021d8aa0
Let --verbose have an argument for the log level.
2024-10-04 10:05:34 -04:00
d854ed0bcf
Allow using SD3 type te output on flux model.
2024-10-03 09:44:54 -04:00
abcd006b8c
Allow more permutations of clip/t5 in dual clip loader.
2024-10-03 09:26:11 -04:00
d985d1d7dc
CLIP Loader node now supports clip_l and clip_g only for SD3.
2024-10-02 04:25:17 -04:00
d1cdf51e1b
Refactor some of the TE detection code.
2024-10-01 07:08:41 -04:00
b4626ab93e
Add simpletuner lycoris format for SD unet.
2024-09-30 06:03:27 -04:00
a9e459c2a4
Use torch.nn.functional.linear in RGB preview code.
...
Add an optional bias to the latent RGB preview code.
2024-09-29 11:27:49 -04:00
3bb4dec720
Fix issue with loras, lowvram and --fast fp8.
2024-09-28 14:42:32 -04:00
8733191563
Flux torch.compile fix ( #5082 )
2024-09-27 22:07:51 -04:00
bdd4a22a2e
Fix flux TE not loading t5 embeddings.
2024-09-24 22:57:22 -04:00
479a427a48
Add dpmpp_2m_cfg_pp ( #4992 )
2024-09-24 02:42:56 -04:00
3a0eeee320
Make --listen listen on both ipv4 and ipv6 at the same time by default.
2024-09-23 04:38:19 -04:00
9c41bc8d10
Remove useless line.
2024-09-23 02:32:29 -04:00
7a415f47a9
Add an optional VAE input to the ControlNetApplyAdvanced node.
...
Deprecate the other controlnet nodes.
2024-09-22 01:24:52 -04:00
dc96a1ae19
Load controlnet in fp8 if weights are in fp8.
2024-09-21 04:50:12 -04:00
2d810b081e
Add load_controlnet_state_dict function.
2024-09-21 01:51:51 -04:00
9f7e9f0547
Add an error message when a controlnet needs a VAE but none is given.
2024-09-21 01:33:18 -04:00
70a708d726
Fix model merging issue.
2024-09-20 02:31:44 -04:00
e7d4782736
add laplace scheduler [2407.03297] ( #4990 )
...
* add laplace scheduler [2407.03297]
* should be here instead lol
* better settings
2024-09-19 23:23:09 -04:00
ad66f7c7d8
Add model_options to load_controlnet function.
2024-09-19 08:23:35 -04:00
de8e8e3b0d
Fix xpu Pytorch nightly build from calling optimize which doesn't exist. ( #4978 )
2024-09-19 05:11:42 -04:00
0b7dfa986d
Improve tiling calculations to reduce number of tiles that need to be processed. ( #4944 )
2024-09-17 03:51:10 -04:00
d514bb38ee
Add some option to model_options for the text encoder.
...
load_device, offload_device and the initial_device can now be set.
2024-09-17 03:49:54 -04:00
0849c80e2a
get_key_patches now works without unloading the model.
2024-09-17 01:57:59 -04:00
e813abbb2c
Long CLIP L support for SDXL, SD3 and Flux.
...
Use the *CLIPLoader nodes.
2024-09-15 07:59:38 -04:00
f48e390032
Support AliMama SD3 and Flux inpaint controlnets.
...
Use the ControlNetInpaintingAliMamaApply node.
2024-09-14 09:05:16 -04:00
cf80d28689
Support loading controlnets with different input.
2024-09-13 09:54:37 -04:00
b962db9952
Add cli arg to override user directory ( #4856 )
...
* Override user directory.
* Use overridden user directory.
* Remove prints.
* Remove references to global user_files.
* Remove unused replace_folder function.
* Remove newline.
* Remove global during get_user_directory.
* Add validation.
2024-09-12 08:10:27 -04:00
9d720187f1
types -> comfy_types to fix import issue.
2024-09-12 03:57:46 -04:00
9f4daca9d9
Doesn't really make sense for cfg_pp sampler to call regular one.
2024-09-11 02:51:36 -04:00
b5d0f2a908
Add CFG++ to DPM++ 2S Ancestral ( #3871 )
...
* Update sampling.py
* Update samplers.py
* my bad
* "fix" the sampler
* Update samplers.py
* i named it wrong
* minor sampling improvements
mainly using a dynamic rho value (hey this sounds a lot like smea!!!)
* revert rho change
rho? r? its just 1/2
2024-09-11 02:49:44 -04:00
9c5fca75f4
Fix lora issue.
2024-09-08 10:10:47 -04:00
32a60a7bac
Support onetrainer text encoder Flux lora.
2024-09-08 09:31:41 -04:00
bb52934ba4
Fix import issue ( #4815 )
2024-09-07 05:28:32 -04:00
ea77750759
Support a generic Comfy format for text encoder loras.
...
This is a format with keys like:
text_encoders.clip_l.transformer.text_model.encoder.layers.9.self_attn.v_proj.lora_up.weight
Instead of waiting for me to add support for specific lora formats you can
convert your text encoder loras to this format instead.
If you want to see an example save a text encoder lora with the SaveLora
node with the commit right after this one.
2024-09-07 02:20:39 -04:00
c27ebeb1c2
Fix onnx export not working on flux.
2024-09-06 03:21:52 -04:00
5cbaa9e07c
Mistoline flux controlnet support.
2024-09-05 00:05:17 -04:00
c7427375ee
Prioritize freeing partially offloaded models first.
2024-09-04 19:47:32 -04:00
f04229b84d
Add emb_patch support to UNetModel forward ( #4779 )
2024-09-04 14:35:15 -04:00
f067ad15d1
Make live preview size a configurable launch argument ( #4649 )
...
* Make live preview size a configurable launch argument
* Remove import from testing phase
* Update cli_args.py
2024-09-03 19:16:38 -04:00
483004dd1d
Support newer glora format.
2024-09-03 17:02:19 -04:00
00a5d08103
Lower fp8 lora memory usage.
2024-09-03 01:25:05 -04:00
d043997d30
Flux onetrainer lora.
2024-09-02 08:22:15 -04:00
8d31a6632f
Speed up inference on nvidia 10 series on Linux.
2024-09-01 17:29:31 -04:00
b643eae08b
Make minimum_inference_memory() depend on --reserve-vram
2024-09-01 01:18:34 -04:00
935ae153e1
Cleanup.
2024-08-30 12:53:59 -04:00
e91662e784
Get logs endpoint & system_stats additions ( #4690 )
...
* Add route for getting output logs
* Include ComfyUI version
* Move to own function
* Changed to memory logger
* Unify logger setup logic
* Fix get version git fallback
---------
Co-authored-by: pythongosssss <125205205+pythongosssss@users.noreply.github.com>
2024-08-30 12:46:37 -04:00
63fafaef45
Fix potential issue with hydit controlnets.
2024-08-30 04:58:41 -04:00
6eb5d64522
Fix glora lowvram issue.
2024-08-29 19:07:23 -04:00
10a79e9898
Implement model part of flux union controlnet.
2024-08-29 18:41:22 -04:00
ea3f39bd69
InstantX depth flux controlnet.
2024-08-29 02:14:19 -04:00
b33cd61070
InstantX canny controlnet.
2024-08-28 19:02:50 -04:00
d31e226650
Unify RMSNorm code.
2024-08-28 16:56:38 -04:00
38c22e631a
Fix case where model was not properly unloaded in merging workflows.
2024-08-27 19:03:51 -04:00
6bbdcd28ae
Support weight padding on diff weight patch ( #4576 )
2024-08-27 13:55:37 -04:00
ab130001a8
Do RMSNorm in native type.
2024-08-27 02:41:56 -04:00
2ca8f6e23d
Make the stochastic fp8 rounding reproducible.
2024-08-26 15:12:06 -04:00
7985ff88b9
Use less memory in float8 lora patching by doing calculations in fp16.
2024-08-26 14:45:58 -04:00
c6812947e9
Fix potential memory leak.
2024-08-26 02:07:32 -04:00
9230f65823
Fix some controlnets OOMing when loading.
2024-08-25 05:54:29 -04:00
8ae23d8e80
Fix onnx export.
2024-08-23 17:52:47 -04:00
7df42b9a23
Fix dora.
2024-08-23 04:58:59 -04:00
5d8bbb7281
Cleanup.
2024-08-23 04:06:27 -04:00
2c1d2375d6
Fix.
2024-08-23 04:04:55 -04:00
64ccb3c7e3
Rework IPEX check for future inclusion of XPU into Pytorch upstream and do a bit more optimization of ipex.optimize(). ( #4562 )
2024-08-23 03:59:57 -04:00
9465b23432
Added SD15_Inpaint_Diffusers model support for unet_config_from_diffusers_unet function ( #4565 )
2024-08-23 03:57:08 -04:00
c0b0da264b
Missing imports.
2024-08-22 17:20:51 -04:00
c26ca27207
Move calculate function to comfy.lora
2024-08-22 17:12:00 -04:00
7c6bb84016
Code cleanups.
2024-08-22 17:05:12 -04:00
c54d3ed5e6
Fix issue with models staying loaded in memory.
2024-08-22 15:58:20 -04:00
c7ee4b37a1
Try to fix some lora issues.
2024-08-22 15:32:18 -04:00
7b70b266d8
Generalize MacOS version check for force-upcast-attention ( #4548 )
...
This code automatically forces upcasting attention for MacOS versions 14.5 and 14.6. My computer returns the string "14.6.1" for `platform.mac_ver()[0]`, so this generalizes the comparison to catch more versions.
I am running MacOS Sonoma 14.6.1 (latest version) and was seeing black image generation on previously functional workflows after recent software updates. This PR solved the issue for me.
See comfyanonymous/ComfyUI#3521
2024-08-22 13:24:21 -04:00
8f60d093ba
Fix issue.
2024-08-22 10:38:24 -04:00
843a7ff70c
fp16 is actually faster than fp32 on a GTX 1080.
2024-08-21 23:23:50 -04:00
a60620dcea
Fix slow performance on 10 series Nvidia GPUs.
2024-08-21 16:39:02 -04:00
015f73dc49
Try a different type of flux fp16 fix.
2024-08-21 16:17:15 -04:00
904bf58e7d
Make --fast work on pytorch nightly.
2024-08-21 14:01:41 -04:00
5f50263088
Replace use of .view with .reshape ( #4522 )
...
When generating images with fp8_e4_m3 Flux and batch size >1, using --fast, ComfyUI throws a "view size is not compatible with input tensor's size and stride" error pointing at the first of these two calls to view.
As reshape is semantically equivalent to view except for working on a broader set of inputs, there should be no downside to changing this. The only difference is that it clones the underlying data in cases where .view would error out. I have confirmed that the output still looks as expected, but cannot confirm that no mutable use is made of the tensors anywhere.
Note that --fast is only marginally faster than the default.
2024-08-21 11:21:48 -04:00
03ec517afb
Remove useless line, adjust windows default reserved vram.
2024-08-21 00:47:19 -04:00
510f3438c1
Speed up fp8 matrix mult by using better code.
2024-08-20 22:53:26 -04:00
ea63b1c092
Simpletrainer lycoris format.
2024-08-20 12:05:13 -04:00
9953f22fce
Add --fast argument to enable experimental optimizations.
...
Optimizations that might break things/lower quality will be put behind
this flag first and might be enabled by default in the future.
Currently the only optimization is float8_e4m3fn matrix multiplication on
4000/ADA series Nvidia cards or later. If you have one of these cards you
will see a speed boost when using fp8_e4m3fn flux for example.
2024-08-20 11:55:51 -04:00
d1a6bd6845
Support loading long clipl model with the CLIP loader node.
2024-08-20 10:46:36 -04:00
83dbac28eb
Properly set if clip text pooled projection instead of using hack.
2024-08-20 10:46:36 -04:00
538cb068bc
Make cast_to a nop if weight is already good.
2024-08-20 10:46:36 -04:00
1b3eee672c
Fix potential issue with multi devices.
2024-08-20 10:46:36 -04:00
9eee470244
New load_text_encoder_state_dicts function.
...
Now you can load text encoders straight from a list of state dicts.
2024-08-19 17:36:35 -04:00
045377ea89
Add a --reserve-vram argument if you don't want comfy to use all of it.
...
--reserve-vram 1.0 for example will make ComfyUI try to keep 1GB vram free.
This can also be useful if workflows are failing because of OOM errors but
in that case please report it if --reserve-vram improves your situation.
2024-08-19 17:16:18 -04:00
4d341b78e8
Bug fixes.
2024-08-19 16:28:55 -04:00
6138f92084
Use better dtype for the lowvram lora system.
2024-08-19 15:35:25 -04:00
be0726c1ed
Remove duplication.
2024-08-19 15:26:50 -04:00
4506ddc86a
Better subnormal fp8 stochastic rounding. Thanks Ashen.
2024-08-19 13:38:03 -04:00
20ace7c853
Code cleanup.
2024-08-19 12:48:59 -04:00
22ec02afc0
Handle subnormal numbers in float8 rounding.
2024-08-19 05:51:08 -04:00
39f114c44b
Less broken non blocking?
2024-08-18 16:53:17 -04:00
6730f3e1a3
Disable non blocking.
...
It fixed some perf issues but caused other issues that need to be debugged.
2024-08-18 14:38:09 -04:00
73332160c8
Enable non blocking transfers in lowvram mode.
2024-08-18 10:29:33 -04:00
2622c55aff
Automatically use RF variant of dpmpp_2s_ancestral if RF model.
2024-08-18 00:47:25 -04:00
1beb348ee2
dpmpp_2s_ancestral_RF for rectified flow (Flux, SD3 and Auraflow).
2024-08-18 00:33:30 -04:00
d31df04c8a
Indentation.
2024-08-17 23:00:44 -04:00
e68763f40c
Add Flux model support for InstantX style controlnet residuals ( #4444 )
...
* Add Flux model support for InstantX style controlnet residuals
* Refactor Flux controlnet residual step to a separate method
* Rollback minor change
* New format for applying controlnet residuals: input->double_blocks, output->single_blocks
* Adjust XLabs Flux controlnet to fit new syntax of applying Flux controlnet residuals
* Remove unnecessary import and minor style change
2024-08-17 22:58:23 -04:00
4f7a3cb6fb
unet -> diffusion_models.
2024-08-17 21:31:04 -04:00
bb222ceddb
Fix loras having a weak effect when applied on fp8.
2024-08-17 15:20:17 -04:00
fca42836f2
Add model_options for text encoder.
2024-08-17 11:17:20 -04:00
cd5017c1c9
calculate_weight function to use a different dtype.
2024-08-17 01:06:08 -04:00
83f343146a
Fix potential lowvram issue.
2024-08-16 17:12:42 -04:00
1770fc77ed
Implement support for taef1 latent previews ( #4409 )
...
* add taef1 handling to several places
* remove guess_latent_channels and add latent_channels info directly to flux model
* remove TODO
* fix numbers
2024-08-16 12:53:13 -04:00
5960f946a9
Move a few files from comfy -> comfy_execution.
...
Python code in the comfy folder should not import things from outside it.
2024-08-15 11:21:14 -04:00
5cfe38f41c
Execution Model Inversion ( #2666 )
...
* Execution Model Inversion
This PR inverts the execution model -- from recursively calling nodes to
using a topological sort of the nodes. This change allows for
modification of the node graph during execution. This allows for two
major advantages:
1. The implementation of lazy evaluation in nodes. For example, if a
"Mix Images" node has a mix factor of exactly 0.0, the second image
input doesn't even need to be evaluated (and visa-versa if the mix
factor is 1.0).
2. Dynamic expansion of nodes. This allows for the creation of dynamic
"node groups". Specifically, custom nodes can return subgraphs that
replace the original node in the graph. This is an incredibly
powerful concept. Using this functionality, it was easy to
implement:
a. Components (a.k.a. node groups)
b. Flow control (i.e. while loops) via tail recursion
c. All-in-one nodes that replicate the WebUI functionality
d. and more
All of those were able to be implemented entirely via custom nodes,
so those features are *not* a part of this PR. (There are some
front-end changes that should occur before that functionality is
made widely available, particularly around variant sockets.)
The custom nodes associated with this PR can be found at:
https://github.com/BadCafeCode/execution-inversion-demo-comfyui
Note that some of them require that variant socket types ("*") be
enabled.
* Allow `input_info` to be of type `None`
* Handle errors (like OOM) more gracefully
* Add a command-line argument to enable variants
This allows the use of nodes that have sockets of type '*' without
applying a patch to the code.
* Fix an overly aggressive assertion.
This could happen when attempting to evaluate `IS_CHANGED` for a node
during the creation of the cache (in order to create the cache key).
* Fix Pyright warnings
* Add execution model unit tests
* Fix issue with unused literals
Behavior should now match the master branch with regard to undeclared
inputs. Undeclared inputs that are socket connections will be used while
undeclared inputs that are literals will be ignored.
* Make custom VALIDATE_INPUTS skip normal validation
Additionally, if `VALIDATE_INPUTS` takes an argument named `input_types`,
that variable will be a dictionary of the socket type of all incoming
connections. If that argument exists, normal socket type validation will
not occur. This removes the last hurdle for enabling variant types
entirely from custom nodes, so I've removed that command-line option.
I've added appropriate unit tests for these changes.
* Fix example in unit test
This wouldn't have caused any issues in the unit test, but it would have
bugged the UI if someone copy+pasted it into their own node pack.
* Use fstrings instead of '%' formatting syntax
* Use custom exception types.
* Display an error for dependency cycles
Previously, dependency cycles that were created during node expansion
would cause the application to quit (due to an uncaught exception). Now,
we'll throw a proper error to the UI. We also make an attempt to 'blame'
the most relevant node in the UI.
* Add docs on when ExecutionBlocker should be used
* Remove unused functionality
* Rename ExecutionResult.SLEEPING to PENDING
* Remove superfluous function parameter
* Pass None for uneval inputs instead of default
This applies to `VALIDATE_INPUTS`, `check_lazy_status`, and lazy values
in evaluation functions.
* Add a test for mixed node expansion
This test ensures that a node that returns a combination of expanded
subgraphs and literal values functions correctly.
* Raise exception for bad get_node calls.
* Minor refactor of IsChangedCache.get
* Refactor `map_node_over_list` function
* Fix ui output for duplicated nodes
* Add documentation on `check_lazy_status`
* Add file for execution model unit tests
* Clean up Javascript code as per review
* Improve documentation
Converted some comments to docstrings as per review
* Add a new unit test for mixed lazy results
This test validates that when an output list is fed to a lazy node, the
node will properly evaluate previous nodes that are needed by any inputs
to the lazy node.
No code in the execution model has been changed. The test already
passes.
* Allow kwargs in VALIDATE_INPUTS functions
When kwargs are used, validation is skipped for all inputs as if they
had been mentioned explicitly.
* List cached nodes in `execution_cached` message
This was previously just bugged in this PR.
2024-08-15 11:21:11 -04:00
0f9c2a7822
Try to fix SDXL OOM issue on some configurations.
2024-08-14 23:08:54 -04:00
f1d6cef71c
Revert "Disable cuda malloc by default."
...
This reverts commit 50bf66e5c4
2024-08-14 08:38:07 -04:00
33fb282d5c
Fix issue.
2024-08-14 02:51:47 -04:00
50bf66e5c4
Disable cuda malloc by default.
2024-08-14 02:49:25 -04:00
a5af64d3ce
Revert "Not sure if this actually changes anything but it can't hurt."
...
This reverts commit 34608de2e9
2024-08-14 01:05:17 -04:00
34608de2e9
Not sure if this actually changes anything but it can't hurt.
2024-08-13 13:29:16 -04:00
39fb74c5bd
Fix bug when model cannot be partially unloaded.
2024-08-13 03:57:55 -04:00
74e124f4d7
Fix some issues with TE being in lowvram mode.
2024-08-12 23:42:21 -04:00
a562c17e8a
load_unet -> load_diffusion_model with a model_options argument.
2024-08-12 23:20:57 -04:00
5942c17d55
Order of operations matters.
2024-08-12 21:56:18 -04:00
c032b11e07
xlabs Flux controlnet implementation. ( #4260 )
...
* xlabs Flux controlnet.
* Fix not working on old python.
* Remove comment.
2024-08-12 21:22:22 -04:00
b8ffb2937f
Memory tweaks.
2024-08-12 15:07:11 -04:00
5d43e75e5b
Fix some issues with the model sometimes not getting patched.
2024-08-12 12:27:54 -04:00
517f4a94e4
Fix some lora loading slowdowns.
2024-08-12 11:50:32 -04:00
52a471c5c7
Change name of log.
2024-08-12 10:35:06 -04:00
ad76574cb8
Fix some potential issues with the previous commits.
2024-08-12 00:23:29 -04:00
9acfe4df41
Support loading directly to vram with CLIPLoader node.
2024-08-12 00:06:01 -04:00
9829b013ea
Fix mistake in last commit.
2024-08-12 00:00:17 -04:00
5c69cde037
Load TE model straight to vram if certain conditions are met.
2024-08-11 23:52:43 -04:00
e9589d6d92
Add a way to set model dtype and ops from load_checkpoint_guess_config.
2024-08-11 08:50:34 -04:00
0d82a798a5
Remove the ckpt_path from load_state_dict_guess_config.
2024-08-11 08:37:35 -04:00
925fff26fd
alternative to `load_checkpoint_guess_config` that accepts a loaded state dict ( #4249 )
...
* make alternative fn
* add back ckpt path as 2nd argument?
2024-08-11 08:36:52 -04:00
75b9b55b22
Fix issues with #4302 and support loading diffusers format flux.
2024-08-10 21:28:24 -04:00
1765f1c60c
FLUX: Added full diffusers mapping for FLUX.1 schnell and dev. Adds full LoRA support from diffusers LoRAs. ( #4302 )
2024-08-10 21:26:41 -04:00
1de69fe4d5
Fix some issues with inference slowing down.
2024-08-10 16:21:25 -04:00
ae197f651b
Speed up hunyuan dit inference a bit.
2024-08-10 07:36:27 -04:00
1b5b8ca81a
Fix regression.
2024-08-09 21:45:21 -04:00
6678d5cf65
Fix regression.
2024-08-09 14:02:38 -04:00
e172564eea
Update controlnet.py to fix the default controlnet weight as constant ( #4285 )
2024-08-09 13:40:05 -04:00
a3cc326748
Better fix for lowvram issue.
2024-08-09 12:16:25 -04:00
86a97e91fc
Fix controlnet regression.
2024-08-09 12:08:58 -04:00
5acdadc9f3
Fix issue with some lowvram weights.
2024-08-09 03:58:28 -04:00
55ad9d5f8c
Fix regression.
2024-08-09 03:36:40 -04:00
a9f04edc58
Implement text encoder part of HunyuanDiT loras.
2024-08-09 03:21:10 -04:00
a475ec2300
Cleanup HunyuanDit controlnets.
...
Use the: ControlNetApply SD3 and HunyuanDiT node.
2024-08-09 02:59:34 -04:00
06eb9fb426
feat: add support for HunYuanDit ControlNet ( #4245 )
...
* add support for HunYuanDit ControlNet
* fix hunyuandit controlnet
* fix typo in hunyuandit controlnet
* fix typo in hunyuandit controlnet
* fix code format style
* add control_weight support for HunyuanDit Controlnet
* use control_weights in HunyuanDit Controlnet
* fix typo
2024-08-09 02:59:24 -04:00
413322645e
Raw torch is faster than einops?
2024-08-08 22:09:29 -04:00
11200de970
Cleaner code.
2024-08-08 20:07:09 -04:00
037c38eb0f
Try to improve inference speed on some machines.
2024-08-08 17:29:27 -04:00
1e11d2d1f5
Better prints.
2024-08-08 17:29:27 -04:00
66d4233210
Fix.
2024-08-08 15:16:51 -04:00
591010b7ef
Support diffusers text attention flux loras.
2024-08-08 14:45:52 -04:00
08f92d55e9
Partial model shift support.
2024-08-08 14:45:06 -04:00
8115d8cce9
Add Flux fp16 support hack.
2024-08-07 15:08:39 -04:00
6969fc9ba4
Make supported_dtypes a priority list.
2024-08-07 15:00:06 -04:00
cb7c4b4be3
Workaround for lora OOM on lowvram mode.
2024-08-07 14:30:54 -04:00
1208863eca
Fix "Comfy" lora keys.
...
They are in this format now:
diffusion_model.full.model.key.name.lora_up.weight
2024-08-07 13:49:31 -04:00
e1c528196e
Fix bundled embed.
2024-08-07 13:30:45 -04:00
17030fd4c0
Support for "Comfy" lora format.
...
The keys are just: model.full.model.key.name.lora_up.weight
It is supported by all comfyui supported models.
Now people can just convert loras to this format instead of having to ask
for me to implement them.
2024-08-07 13:18:32 -04:00
c19dcd362f
Controlnet code refactor.
2024-08-07 12:59:28 -04:00
1c08bf35b4
Support format for embeddings bundled in loras.
2024-08-07 03:45:25 -04:00
b334605a66
Fix OOMs happening in some cases.
...
A cloned model patcher sometimes reported a model was loaded on a device
when it wasn't.
2024-08-06 13:36:04 -04:00
c14ac98fed
Unload models and load them back in lowvram mode no free vram.
2024-08-06 03:22:39 -04:00
2d75df45e6
Flux tweak memory usage.
2024-08-05 21:58:28 -04:00
8edbcf5209
Improve performance on some lowend GPUs.
2024-08-05 16:24:04 -04:00
a178e25912
Fix Flux FP64 math on XPU ( #4210 )
2024-08-05 01:26:20 -04:00
78e133d041
Support simple diffusers Flux loras.
2024-08-04 22:05:48 -04:00
7afa985fba
Correct spelling 'token_weight_pars_t5' to 'token_weight_pairs_t5' ( #4200 )
2024-08-04 17:10:02 -04:00
3b71f84b50
ONNX tracing fixes.
2024-08-04 15:45:43 -04:00
0a6b008117
Fix issue with some custom nodes.
2024-08-04 10:03:33 -04:00
f7a5107784
Fix crash.
2024-08-03 16:55:38 -04:00
91be9c2867
Tweak lowvram memory formula.
2024-08-03 16:44:50 -04:00
03c5018c98
Lower lowvram memory to 1/3 of free memory.
2024-08-03 15:14:07 -04:00
2ba5cc8b86
Fix some issues.
2024-08-03 15:06:40 -04:00
1e68002b87
Cap lowvram to half of free memory.
2024-08-03 14:50:20 -04:00
ba9095e5bd
Automatically use fp8 for diffusion model weights if:
...
Checkpoint contains weights in fp8.
There isn't enough memory to load the diffusion model in GPU vram.
2024-08-03 13:45:19 -04:00
f123328b82
Load T5 in fp8 if it's in fp8 in the Flux checkpoint.
2024-08-03 12:39:33 -04:00
63a7e8edba
More aggressive batch splitting.
2024-08-03 11:53:30 -04:00
ea03c9dcd2
Better per model memory usage estimations.
2024-08-02 18:09:24 -04:00
3a9ee995cf
Tweak regular SD memory formula.
2024-08-02 17:34:30 -04:00
47da42d928
Better Flux vram estimation.
2024-08-02 17:02:35 -04:00
ce9ac2fe05
Fix clip_g/clip_l mixup ( #4168 )
2024-08-01 21:40:56 -04:00
e638f2858a
Hack to make all resolutions work on Flux models.
2024-08-01 21:39:18 -04:00
d420bc792a
Tweak the memory usage formulas for Flux and SD.
2024-08-01 17:53:45 -04:00
d965474aaa
Make ComfyUI split batches a higher priority than weight offload.
2024-08-01 16:39:59 -04:00
1c61361fd2
Fast preview support for Flux.
2024-08-01 16:28:11 -04:00
a6decf1e62
Fix bfloat16 potentially not being enabled on mps.
2024-08-01 16:18:44 -04:00
48eb1399c0
Try to fix mac issue.
2024-08-01 13:41:27 -04:00
d7430a1651
Add a way to load the diffusion model in fp8 with UNETLoader node.
2024-08-01 13:30:51 -04:00
f2b80f95d2
Better Mac support on flux model.
2024-08-01 13:10:50 -04:00
1aa9cf3292
Make lowvram more aggressive on low memory machines.
2024-08-01 12:11:57 -04:00
eb96c3bd82
Fix .sft file loading (they are safetensors files).
2024-08-01 11:32:58 -04:00
5f98de7697
Load flux t5 in fp8 if weights are in fp8.
2024-08-01 11:05:56 -04:00
8d34211a7a
Fix old python versions no longer working.
2024-08-01 09:57:20 -04:00
1589b58d3e
Basic Flux Schnell and Flux Dev model implementation.
2024-08-01 09:49:29 -04:00
7ad574bffd
Mac supports bf16 just make sure you are using the latest pytorch.
2024-08-01 09:42:17 -04:00
e2382b6adb
Make lowvram less aggressive when there are large amounts of free memory.
2024-08-01 03:58:58 -04:00
c24f897352
Fix to get fp8 working on T5 base.
2024-07-31 02:00:19 -04:00
a5991a7aa6
Fix hunyuan dit text encoder weights always being in fp32.
2024-07-31 01:34:57 -04:00
2c038ccef0
Lower CLIP memory usage by a bit.
2024-07-31 01:32:35 -04:00
b85216a3c0
Lower T5 memory usage by a few hundred MB.
2024-07-31 00:52:34 -04:00
82cae45d44
Fix potential issue with non clip text embeddings.
2024-07-30 14:41:13 -04:00
25853d0be8
Use common function for casting weights to input.
2024-07-30 10:49:14 -04:00
79040635da
Remove unnecessary code.
2024-07-30 05:01:34 -04:00
66d35c07ce
Improve artifacts on hydit, auraflow and SD3 on specific resolutions.
...
This breaks seeds for resolutions that are not a multiple of 16 in pixel
resolution by using circular padding instead of reflection padding but
should lower the amount of artifacts when doing img2img at those
resolutions.
2024-07-29 20:48:50 -04:00
4ba7fa0244
Refactor: Move sd2_clip.py to text_encoders folder.
2024-07-28 01:19:20 -04:00
cf4418b806
Don't treat Bert model like CLIP.
...
Bert can accept up to 512 tokens so any prompt with more than 77 should
just be passed to it as is instead of splitting it up like CLIP.
2024-07-26 13:08:12 -04:00
8328a2d8cd
Let hunyuan dit work with all prompt lengths.
2024-07-26 12:11:32 -04:00
afe732bef9
Hunyuan dit can now accept longer prompts.
2024-07-26 11:52:58 -04:00
a9ac56fc0d
Own BertModel implementation that works with lowvram.
2024-07-26 04:47:17 -04:00
25b51b1a8b
Hunyuan DiT lora support.
2024-07-25 22:42:54 -04:00
a5f4292f9f
Basic hunyuan dit implementation. ( #4102 )
...
* Let tokenizers return weights to be stored in the saved checkpoint.
* Basic hunyuan dit implementation.
* Fix some resolutions not working.
* Support hydit checkpoint save.
* Init with right dtype.
* Switch to optimized attention in pooler.
* Fix black images on hunyuan dit.
2024-07-25 18:21:08 -04:00
f87810cd3e
Let tokenizers return weights to be stored in the saved checkpoint.
2024-07-25 10:52:09 -04:00
10c919f4c7
Make it possible to load tokenizer data from checkpoints.
2024-07-24 16:43:53 -04:00
10b43ceea5
Remove duplicate code.
2024-07-24 01:12:59 -04:00
0a4c49c57c
Support MT5.
2024-07-23 15:35:28 -04:00
88ed893034
Allow SPieceTokenizer to load model from a byte string.
2024-07-23 14:17:42 -04:00
334ba48cea
More generic unet prefix detection code.
2024-07-23 14:13:32 -04:00
14764aa2e2
Rename LLAMATokenizer to SPieceTokenizer.
2024-07-22 12:21:45 -04:00
b2c995f623
"auto" type is only relevant to the SetUnionControlNetType node.
2024-07-22 11:30:38 -04:00
4151fbfa8a
Add error message on union controlnet ( #4081 )
2024-07-22 11:27:32 -04:00
95fa9545f1
Only append zero to noise schedule if last sigma isn't zero.
2024-07-20 12:37:30 -04:00
6ab8cad22e
Implement beta sampling scheduler.
...
It is based on: https://arxiv.org/abs/2407.12173
Add "beta" to the list of schedulers and the BetaSamplingScheduler node.
2024-07-19 18:05:09 -04:00
855789403b
support clip-vit-large-patch14-336 ( #4042 )
...
* support clip-vit-large-patch14-336
* support clip-vit-large-patch14-336
2024-07-17 13:12:50 -04:00
6f7869f365
Get clip vision image size from config.
2024-07-17 13:05:38 -04:00
281ad42df4
Fix lowvram union controlnet bug.
2024-07-17 10:16:31 -04:00
c5a48b15bd
Make default hash lib configurable without code changes via CLI argument ( #3947 )
...
* cli_args: Add --duplicate-check-hash-function.
* server.py: compare_image_hash configurable hash function
Uses an argument added in cli_args to specify the type of hashing to default to for duplicate hash checking. Uses an `eval()` to identify the specific hashlib class to utilize, but ultimately safely operates because we have specific options and only those options/choices in the arg parser. So we don't have any unsafe input there.
* Add hasher() to node_helpers
* hashlib selection moved to node_helpers
* default-hashing-function instead of dupe checking hasher
This makes a default-hashing-function option instead of previous selected option.
* Use args.default_hashing_function
* Use safer handling for node_helpers.hasher()
Uses a safer handling method than `eval` to evaluate default hashing function.
* Stray parentheses are evil.
* Indentation fix.
Somehow when I hit save I didn't notice I missed a space to make indentation work proper. Oops!
2024-07-16 18:27:09 -04:00
8270c62530
Add SetUnionControlNetType to set the type of the union controlnet model.
2024-07-16 17:04:53 -04:00
821f93872e
Allow model sampling to set number of timesteps.
2024-07-16 15:18:40 -04:00
99458e8aca
Add `FrontendManager` to manage non-default front-end impl ( #3897 )
...
* Add frontend manager
* Add tests
* nit
* Add unit test to github CI
* Fix path
* nit
* ignore
* Add logging
* Install test deps
* Remove 'stable' keyword support
* Update test
* Add web-root arg
* Rename web-root to front-end-root
* Add test on non-exist version number
* Use repo owner/name to replace hard coded provider list
* Inline cmd args
* nit
* Fix unit test
2024-07-16 11:26:11 -04:00
1305fb294c
Refactor: Move some code to the comfy/text_encoders folder.
2024-07-15 17:36:24 -04:00
7914c47d5a
Quick fix for the promax controlnet.
2024-07-14 10:07:36 -04:00
a3dffc447a
Support AuraFlow Lora and loading model weights in diffusers format.
...
You can load model weights in diffusers format using the UNETLoader node.
2024-07-13 13:51:40 -04:00
29c2e26724
Better tokenizing code for AuraFlow.
2024-07-12 01:15:25 -04:00
8e012043a9
Add a ModelSamplingAuraFlow node to change the shift value.
...
Set the default AuraFlow shift value to 1.73 (sqrt(3)).
2024-07-11 17:57:36 -04:00
9f291d75b3
AuraFlow model implementation.
2024-07-11 16:52:26 -04:00
f45157e3ac
Fix error message never being shown.
2024-07-11 11:46:51 -04:00
5e1fced639
Cleaner support for loading different diffusion model types.
2024-07-11 11:37:31 -04:00
ffe0bb0a33
Remove useless code.
2024-07-10 20:33:12 -04:00
391c1046cf
More flexibility with text encoder return values.
...
Text encoders can now return other values to the CONDITIONING than the cond
and pooled output.
2024-07-10 20:06:50 -04:00
e44fa5667f
Support returning text encoder attention masks.
2024-07-10 19:31:22 -04:00
f1a01c2c7e
Add sampler_pre_cfg_function ( #3979 )
...
* Update samplers.py
* Update model_patcher.py
2024-07-09 16:20:49 -04:00
ade7aa1b0c
Remove useless import.
2024-07-09 11:05:05 -04:00
faa57430b0
Controlnet union model basic implementation.
...
This is only the model code itself, it currently defaults to an empty
embedding [0] * 6 which seems to work better than treating it like a
regular controlnet.
TODO: Add nodes to select the image type.
2024-07-08 23:49:02 -04:00
bb663bcd6c
Rename clip_t5base to t5base for stable audio text encoder.
2024-07-08 08:53:55 -04:00
2dc84d1444
Add a way to set the timestep multiplier in the flow sampling.
2024-07-06 04:06:03 -04:00
ff63893d10
Support other types of T5 models.
2024-07-06 02:42:53 -04:00
4040491149
Better T5xxl detection.
2024-07-06 00:53:33 -04:00
b8e58a9394
Cleanup T5 code a bit.
2024-07-06 00:36:49 -04:00
80c4590998
Allow specifying the padding token for the tokenizer.
2024-07-06 00:06:49 -04:00
ce649d61c0
Allow zeroing out of embeds with unused attention mask.
2024-07-05 23:48:17 -04:00
739b76630e
Remove useless code.
2024-07-04 15:14:13 -04:00
d7484ef30c
Support loading checkpoints with the UNETLoader node.
2024-07-03 11:34:32 -04:00
537f35c7bc
Don't update dict if contiguous.
2024-07-02 20:21:51 -04:00
3f46362d22
fix non-contiguous tensor saving (from channels-last) ( #3932 )
2024-07-02 20:16:33 -04:00
9dd549e253
Add `--no-custom-node` cmd flag ( #3903 )
...
* Add --no-custom-node cmd flag
* nit
2024-07-01 17:54:03 -04:00
05e831697a
Switch to the real cfg++ method in the samplers.
...
The old _pp ones will be updated automatically to the regular ones with 2x
the cfg.
My fault for not checking what the "_pp" samplers actually did.
2024-06-29 11:59:48 -04:00
264caca20e
ControlNetApplySD3 node can now be used to use SD3 controlnets.
2024-06-27 18:43:11 -04:00
f8f7568d03
Basic SD3 controlnet implementation.
...
Still missing the node to properly use it.
2024-06-27 18:43:11 -04:00
66aaa14001
Controlnet refactor.
2024-06-27 18:43:11 -04:00
8ceb5a02a3
Support saving stable audio checkpoint that can be loaded back.
2024-06-27 11:06:52 -04:00
4f9d2b057c
Remove print.
2024-06-27 02:54:15 -04:00
44947e7ad4
Add DEIS order 3 sampler.
...
Order 4 seems to give bad results.
2024-06-26 22:40:05 -04:00
69d710e40f
Implement my alternative take on CFG++ as the euler_pp sampler.
...
Add euler_ancestral_pp which is the ancestral version of euler with the
same modification.
2024-06-25 07:41:52 -04:00
73ca780019
Add SamplerEulerCFG++ node.
...
This node should match the DDIM implementation of CFG++ when "regular" is
selected.
"alternative" is a slightly different take on CFG++
2024-06-23 13:21:18 -04:00
2f360ae898
Support OneTrainer SD3 lora format.
2024-06-22 13:08:04 -04:00
4ef1479dcd
Multi dimension tiled scale function and tiled VAE audio encoding fallback.
2024-06-22 11:57:49 -04:00
1e2839f4d9
More proper tiled audio decoding.
2024-06-20 16:50:31 -04:00
d5efde89b7
Add ipndm_v sampler, works best with the exponential scheduler.
2024-06-20 08:51:49 -04:00
028a583bef
Fix issue with full diffusers SD3 loras.
2024-06-19 22:32:04 -04:00
0d6a57938e
Support loading diffusers SD3 model format with UNETLoader node.
2024-06-19 22:21:18 -04:00
b08a9dd04b
Remove empty line.
2024-06-19 20:20:35 -04:00
eee815ec99
Update sd1_clip.py ( #3684 )
...
Made token instance check more flexible so it also works with integers from numpy arrays or long tensors
2024-06-19 16:42:41 -04:00
e11052afcf
Add ipndm sampler.
2024-06-19 16:32:30 -04:00
3914d5a2ae
Support full SD3 loras.
2024-06-19 10:13:33 -04:00
a45df69570
Basic tiled decoding for audio VAE.
2024-06-17 22:48:23 -04:00
b7c473d1ab
Fix lora keys for SimpleTuner ( #3759 )
2024-06-17 07:55:06 -04:00
6425252c4f
Use fp16 as the default vae dtype for the audio VAE.
2024-06-16 13:12:54 -04:00
8ddc151a4c
Squash depreciation warning on new pytorch.
2024-06-16 13:06:23 -04:00
ca9d300a80
Better estimation for memory usage during audio VAE encoding/decoding.
2024-06-16 11:47:32 -04:00
746a0410d4
Fix VAEEncode with taesd3.
2024-06-16 03:10:04 -04:00
04e8798c37
Improvements to the TAESD3 implementation.
2024-06-16 02:04:24 -04:00
df7db0e027
support TAESD3 ( #3738 )
2024-06-16 02:03:53 -04:00
bb1969cab7
Initial support for the stable audio open model.
2024-06-15 12:14:56 -04:00
1281f933c1
Small optimization.
2024-06-15 02:44:38 -04:00
f2e844e054
Optimize some unneeded if conditions in the sampling code.
2024-06-15 02:26:19 -04:00
0ec513d877
Add a --force-channels-last to inference models in channel last mode.
2024-06-15 01:08:12 -04:00
0e06b370db
Print key names for easier debugging.
2024-06-14 18:18:53 -04:00
5eb98f0092
Exempt IPEX from non_blocking previews fixing segmentation faults. ( #3708 )
2024-06-13 18:51:14 -04:00
ac151ac169
Support SD3 diffusers lora.
2024-06-13 18:26:10 -04:00
37a08a41b3
Support setting weight offsets in weight patcher.
2024-06-13 17:21:26 -04:00
605e64f6d3
Fix lowvram issue.
2024-06-12 10:39:33 -04:00
1ddf512fdc
Don't auto convert clip and vae weights to fp16 when saving checkpoint.
2024-06-12 01:07:58 -04:00
694e0b48e0
SD3 better memory usage estimation.
2024-06-12 00:49:00 -04:00
69c8d6d8a6
Single and dual clip loader nodes support SD3.
...
You can use the CLIPLoader to use the t5xxl only or the DualCLIPLoader to
use CLIP-L and CLIP-G only for sd3.
2024-06-11 23:27:39 -04:00
0e49211a11
Load the SD3 T5xxl model in the same dtype stored in the checkpoint.
2024-06-11 17:03:26 -04:00
5889b7ca0a
Support multiple text encoder configurations on SD3.
2024-06-11 13:14:43 -04:00
9424522ead
Reuse code.
2024-06-11 07:20:26 -04:00
73ce178021
Remove redundancy in mmdit.py ( #3685 )
2024-06-11 06:30:25 -04:00
a82fae2375
Fix bug with cosxl edit model.
2024-06-10 16:00:03 -04:00
8c4a9befa7
SD3 Support.
2024-06-10 14:06:23 -04:00
a5e6a632f9
Support sampling non 2D latents.
2024-06-10 01:31:09 -04:00
742d5720d1
Support zeroing out text embeddings with the attention mask.
2024-06-09 16:51:58 -04:00
6cd8ffc465
Reshape the empty latent image to the right amount of channels if needed.
2024-06-08 02:35:08 -04:00
56333d4850
Use the end token for the text encoder attention mask.
2024-06-07 03:05:23 -04:00
104fcea0c8
Add function to get the list of currently loaded models.
2024-06-05 23:25:16 -04:00
b1fd26fe9e
pytorch xpu should be flash or mem efficient attention?
2024-06-04 17:44:14 -04:00
809cc85a8e
Remove useless code.
2024-06-02 19:23:37 -04:00
b249862080
Add an annoying print to a function I want to remove.
2024-06-01 12:47:31 -04:00
bf3e334d46
Disable non_blocking when --deterministic or directml.
2024-05-30 11:07:38 -04:00
b26da2245f
Fix UnetParams annotation typo ( #3589 )
2024-05-27 19:30:35 -04:00
0920e0e5fe
Remove some unused imports.
2024-05-27 19:08:27 -04:00
ffc4b7c30e
Fix DORA strength.
...
This is a different version of #3298 with more correct behavior.
2024-05-25 02:50:11 -04:00
efa5a711b2
Reduce memory usage when applying DORA: #3557
2024-05-24 23:36:48 -04:00
6c23854f54
Fix OSX latent2rgb previews.
2024-05-22 13:56:28 -04:00
7718ada4ed
Add type annotation UnetWrapperFunction ( #3531 )
...
* Add type annotation UnetWrapperFunction
* nit
* Add types.py
2024-05-22 02:07:27 -04:00
8508df2569
Work around black image bug on Mac 14.5 by forcing attention upcasting.
2024-05-21 16:56:33 -04:00
83d969e397
Disable xformers when tracing model.
2024-05-21 13:55:49 -04:00
1900e5119f
Fix potential issue.
2024-05-20 08:19:54 -04:00
09e069ae6c
Log the pytorch version.
2024-05-20 06:22:29 -04:00
11a2ad5110
Fix controlnet not upcasting on models that have it enabled.
2024-05-19 17:58:03 -04:00
0bdc2b15c7
Cleanup.
2024-05-18 10:11:44 -04:00
98f828fad9
Remove unnecessary code.
2024-05-18 09:36:44 -04:00
19300655dd
Don't automatically switch to lowvram mode on GPUs with low memory.
2024-05-17 00:31:32 -04:00
46daf0a9a7
Add debug options to force on and off attention upcasting.
2024-05-16 04:09:41 -04:00
2d41642716
Fix lowvram dora issue.
2024-05-15 02:47:40 -04:00
ec6f16adb6
Fix SAG.
2024-05-14 18:02:27 -04:00
bb4940d837
Only enable attention upcasting on models that actually need it.
2024-05-14 17:00:50 -04:00
b0ab31d06c
Refactor attention upcasting code part 1.
2024-05-14 12:47:31 -04:00
f509c6fe21
Fix Intel GPU memory allocation accuracy and documentation update. ( #3459 )
...
* Change calculation of memory total to be more accurate, allocated is actually smaller than reserved.
* Update README.md install documentation for Intel GPUs.
2024-05-12 06:36:30 -04:00
fa6dd7e5bb
Fix lowvram issue with saving checkpoints.
...
The previous fix didn't cover the case where the model was loaded in
lowvram mode right before.
2024-05-12 06:13:45 -04:00
49c20cdc70
No longer necessary.
2024-05-12 05:34:43 -04:00
e1489ad257
Fix issue with lowvram mode breaking model saving.
2024-05-11 21:55:20 -04:00
93e876a3be
Remove warnings that confuse people.
2024-05-09 05:29:42 -04:00
cd07340d96
Typo fix.
2024-05-08 18:36:56 -04:00
c61eadf69a
Make the load checkpoint with config function call the regular one.
...
I was going to completely remove this function because it is unmaintainable
but I think this is the best compromise.
The clip skip and v_prediction parts of the configs should still work but
not the fp16 vs fp32.
2024-05-06 20:04:39 -04:00
a56d02efc7
Change torch.xpu to ipex.optimize, xpu device initialization and remove workaround for text node issue from older IPEX. ( #3388 )
2024-05-02 03:26:50 -04:00
f81a6fade8
Fix some edge cases with samplers and arrays with a single sigma.
2024-05-01 17:05:30 -04:00
2aed53c4ac
Workaround xformers bug.
2024-04-30 21:23:40 -04:00
bacce529fb
Add TLS Support ( #3312 )
...
* Add TLS Support
* Add to readme
* Add guidance for windows users on generating certificates
* Add guidance for windows users on generating certificates
* Fix typo
2024-04-30 20:17:02 -04:00
7990ae18c1
Fix error when more cond masks passed in than batch size ( #3353 )
2024-04-26 12:51:12 -04:00
8dc19e40d1
Don't init a VAE model when there are no VAE weights.
2024-04-24 09:20:31 -04:00
c59fe9f254
Support VAE without quant_conv.
2024-04-18 21:05:33 -04:00
719fb2c81d
Add basic PAG node.
2024-04-14 23:49:50 -04:00
258dbc06c3
Fix some memory related issues.
2024-04-14 12:08:58 -04:00
58812ab8ca
Support SDXS 512 model.
2024-04-12 22:12:35 -04:00
831511a1ee
Fix issue with sampling_settings persisting across models.
2024-04-09 23:20:43 -04:00
30abc324c2
Support properly saving CosXL checkpoints.
2024-04-08 00:36:22 -04:00
0a03009808
Fix issue with controlnet models getting loaded multiple times.
2024-04-06 18:38:39 -04:00
38ed2da2dd
Fix typo in lowvram patcher ( #3209 )
2024-04-05 12:02:13 -04:00
1088d1850f
Support for CosXL models.
2024-04-05 10:53:41 -04:00
41ed7e85ea
Fix object_patches_backup not being the same object across clones.
2024-04-05 00:22:44 -04:00
0f5768e038
Fix missing arguments in cfg_function.
2024-04-04 23:38:57 -04:00
1f4fc9ea0c
Fix issue with get_model_object on patched model.
2024-04-04 23:01:02 -04:00
1a0486bb96
Fix model needing to be loaded on GPU to generate the sigmas.
2024-04-04 22:08:49 -04:00
c6bd456c45
Make zero denoise a NOP.
2024-04-04 11:41:27 -04:00
fcfd2bdf8a
Small cleanup.
2024-04-04 11:16:49 -04:00
0542088ef8
Refactor sampler code for more advanced sampler nodes part 2.
2024-04-04 01:26:41 -04:00
57753c964a
Refactor sampling code for more advanced sampler nodes.
2024-04-03 22:09:51 -04:00
6c6a39251f
Fix saving text encoder in fp8.
2024-04-02 11:46:34 -04:00
e6482fbbfc
Refactor calc_cond_uncond_batch into calc_cond_batch.
...
calc_cond_batch can take an arbitrary amount of cond inputs.
Added a calc_cond_uncond_batch wrapper with a warning so custom nodes
won't break.
2024-04-01 18:07:47 -04:00
575acb69e4
IP2P model loading support.
...
This is the code to load the model and inference it with only a text
prompt. This commit does not contain the nodes to properly use it with an
image input.
This supports both the original SD1 instructpix2pix model and the
diffusers SDXL one.
2024-03-31 03:10:28 -04:00
94a5a67c32
Cleanup to support different types of inpaint models.
2024-03-29 14:44:13 -04:00
5d8898c056
Fix some performance issues with weight loading and unloading.
...
Lower peak memory usage when changing model.
Fix case where model weights would be unloaded and reloaded.
2024-03-28 18:04:42 -04:00
327ca1313d
Support SDXS 0.9
2024-03-27 23:58:58 -04:00
ae77590b4e
dora_scale support for lora file.
2024-03-25 18:09:23 -04:00
c6de09b02e
Optimize memory unload strategy for more optimized performance.
2024-03-24 02:36:30 -04:00
0624838237
Add inverse noise scaling function.
2024-03-21 14:49:11 -04:00
5d875d77fe
Fix regression with lcm not working with batches.
2024-03-20 20:48:54 -04:00
4b9005e949
Fix regression with model merging.
2024-03-20 13:56:12 -04:00
c18a203a8a
Don't unload model weights for non weight patches.
2024-03-20 02:27:58 -04:00
150a3e946f
Make LCM sampler use the model noise scaling function.
2024-03-20 01:35:59 -04:00
40e124c6be
SV3D support.
2024-03-18 16:54:13 -04:00
cacb022c4a
Make saved SD1 checkpoints match more closely the official one.
2024-03-18 00:26:23 -04:00
d7897fff2c
Move cascade scale factor from stage_a to latent_formats.py
2024-03-16 14:49:35 -04:00
f2fe635c9f
SamplerDPMAdaptative node to test the different options.
2024-03-15 22:36:10 -04:00
448d9263a2
Fix control loras breaking.
2024-03-14 09:30:21 -04:00
db8b59ecff
Lower memory usage for loras in lowvram mode at the cost of perf.
2024-03-13 20:07:27 -04:00
2a813c3b09
Switch some more prints to logging.
2024-03-11 16:34:58 -04:00
0ed72befe1
Change log levels.
...
Logging level now defaults to info. --verbose sets it to debug.
2024-03-11 13:54:56 -04:00
65397ce601
Replace prints with logging and add --verbose argument.
2024-03-10 12:14:23 -04:00
5f60ee246e
Support loading the sr cascade controlnet.
2024-03-07 01:22:48 -05:00
03e6e81629
Set upscale algorithm to bilinear for stable cascade controlnet.
2024-03-06 02:59:40 -05:00
03e83bb5d0
Support stable cascade canny controlnet.
2024-03-06 02:25:42 -05:00
10860bcd28
Add compression_ratio to controlnet code.
2024-03-05 15:15:20 -05:00
478f71a249
Remove useless check.
2024-03-04 08:51:25 -05:00
12c1080ebc
Simplify differential diffusion code.
2024-03-03 15:34:42 -05:00
727021bdea
Implement Differential Diffusion ( #2876 )
...
* Implement Differential Diffusion
* Cleanup.
* Fix.
* Masks should be applied at full strength.
* Fix colors.
* Register the node.
* Cleaner code.
* Fix issue with getting unipc sampler.
* Adjust thresholds.
* Switch to linear thresholds.
* Only calculate nearest_idx on valid thresholds.
2024-03-03 15:34:13 -05:00
1abf8374ec
utils.set_attr can now be used to set any attribute.
...
The old set_attr has been renamed to set_attr_param.
2024-03-02 17:27:23 -05:00
dce3555339
Add some tesla pascal GPUs to the fp16 working but slower list.
2024-03-02 17:16:31 -05:00
51df846598
Let conditioning specify custom concat conds.
2024-03-02 11:44:06 -05:00
9f71e4b62d
Let model patches patch sub objects.
2024-03-02 11:43:27 -05:00
00425563c0
Cleanup: Use sampling noise scaling function for inpainting.
2024-03-01 14:24:41 -05:00
c62e836167
Move noise scaling to object with sampling math.
2024-03-01 12:54:38 -05:00
cb7c3a2921
Allow image_only_indicator to be None.
2024-02-29 13:11:30 -05:00
b3e97fc714
Koala 700M and 1B support.
...
Use the UNET Loader node to load the unet file to use them.
2024-02-28 12:10:11 -05:00
37a86e4618
Remove duplicate text_projection key from some saved models.
2024-02-28 03:57:41 -05:00
8daedc5bf2
Auto detect playground v2.5 model.
2024-02-27 18:03:03 -05:00
d46583ecec
Playground V2.5 support with ModelSamplingContinuousEDM node.
...
Use ModelSamplingContinuousEDM with edm_playground_v2.5 selected.
2024-02-27 15:12:33 -05:00
1e0fcc9a65
Make XL checkpoints save in a more standard format.
2024-02-27 02:07:40 -05:00
b416be7d78
Make the text projection saved in the checkpoint the right format.
2024-02-27 01:52:23 -05:00
03c47fc0f2
Add a min_length property to tokenizer class.
2024-02-26 21:36:37 -05:00
8ac69f62e5
Make return_projected_pooled setable from the __init__
2024-02-25 14:49:13 -05:00
ca7c310a0e
Support loading old CLIP models saved with CLIPSave.
2024-02-25 08:29:12 -05:00
c2cb8e889b
Always return unprojected pooled output for gligen.
2024-02-25 07:33:13 -05:00
1cb3f6a83b
Move text projection into the CLIP model code.
...
Fix issue with not loading the SSD1B clip correctly.
2024-02-25 01:41:08 -05:00
6533b172c1
Support text encoder text_projection in lora.
2024-02-24 23:50:46 -05:00
1e5f0f66be
Support lora keys with lora_prior_unet_ and lora_prior_te_
2024-02-23 12:21:20 -05:00
e1cb93c383
Fix model and cond transformer options merge
2024-02-23 01:19:43 -07:00
10847dfafe
Cleanup uni_pc inpainting.
...
This causes some small changes to the uni pc inpainting behavior but it
seems to improve results slightly.
2024-02-23 02:39:35 -05:00
18c151b3e3
Add some latent2rgb matrices for previews.
2024-02-20 10:57:24 -05:00
0d0fbabd1d
Pass pooled CLIP to stage b.
2024-02-20 04:24:45 -05:00
c6b7a157ed
Align simple scheduling closer to official stable cascade scheduler.
2024-02-20 04:24:39 -05:00
88f300401c
Enable fp16 by default on mps.
2024-02-19 12:00:48 -05:00
e93cdd0ad0
Remove print.
2024-02-19 11:47:26 -05:00
3711b31dff
Support Stable Cascade in checkpoint format.
2024-02-19 11:20:48 -05:00
d91f45ef28
Some cleanups to how the text encoders are loaded.
2024-02-19 10:46:30 -05:00
a7b5eaa7e3
Forgot to commit this.
2024-02-19 04:25:46 -05:00
3b2e579926
Support loading the Stable Cascade effnet and previewer as a VAE.
...
The effnet can be used to encode images for img2img with Stage C.
2024-02-19 04:10:01 -05:00
dccca1daa5
Fix gligen lowvram mode.
2024-02-18 02:20:23 -05:00
8b60d33bb7
Add ModelSamplingStableCascade to control the shift sampling parameter.
...
shift is 2.0 by default on Stage C and 1.0 by default on Stage B.
2024-02-18 00:55:23 -05:00
6bcf57ff10
Fix attention masks properly for multiple batches.
2024-02-17 16:15:18 -05:00
11e3221f1f
fp8 weight support for Stable Cascade.
2024-02-17 15:27:31 -05:00
f8706546f3
Fix attention mask batch size in some attention functions.
2024-02-17 15:22:21 -05:00
3b9969c1c5
Properly fix attention masks in CLIP with batches.
2024-02-17 12:13:13 -05:00
5b40e7a5ed
Implement shift schedule for cascade stage C.
2024-02-17 11:38:47 -05:00
929e266f3e
Manual cast for bf16 on older GPUs.
2024-02-17 09:01:17 -05:00
6c875d846b
Fix clip attention mask issues on some hardware.
2024-02-17 07:53:52 -05:00
805c36ac9c
Make Stable Cascade work on old pytorch 2.0
2024-02-17 00:42:30 -05:00
f2d1d16f4f
Support Stable Cascade Stage B lite.
2024-02-16 23:41:23 -05:00
0b3c50480c
Make --force-fp32 disable loading models in bf16.
2024-02-16 23:01:54 -05:00
97d03ae04a
StableCascade CLIP model support.
2024-02-16 13:29:04 -05:00
667c92814e
Stable Cascade Stage B.
2024-02-16 13:02:03 -05:00
f83109f09b
Stable Cascade Stage C.
2024-02-16 10:55:08 -05:00
5e06baf112
Stable Cascade Stage A.
2024-02-16 06:30:39 -05:00
aeaeca10bd
Small refactor of is_device_* functions.
2024-02-15 21:10:10 -05:00
38b7ac6e26
Don't init the CLIP model when the checkpoint has no CLIP weights.
2024-02-13 00:01:08 -05:00
7dd352cbd7
Merge branch 'feature_expose_discard_penultimate_sigma' of https://github.com/blepping/ComfyUI
2024-02-11 12:23:30 -05:00
f44225fd5f
Fix infinite while loop being possible in ddim_scheduler
2024-02-09 17:11:34 -06:00
25a4805e51
Add a way to set different conditioning for the controlnet.
2024-02-09 14:13:31 -05:00
a352c021ec
Allow custom samplers to request discard penultimate sigma
2024-02-08 02:24:23 -07:00
c661a8b118
Don't use numpy for calculating sigmas.
2024-02-07 18:52:51 -05:00
236bda2683
Make minimum tile size the size of the overlap.
2024-02-05 01:29:26 -05:00
66e28ef45c
Don't use is_bf16_supported to check for fp16 support.
2024-02-04 20:53:35 -05:00
24129d78e6
Speed up SDXL on 16xx series with fp16 weights and manual cast.
2024-02-04 13:23:43 -05:00
4b0239066d
Always use fp16 for the text encoders.
2024-02-02 10:02:49 -05:00
da7a8df0d2
Put VAE key name in model config.
2024-01-30 02:24:38 -05:00
89507f8adf
Remove some unused imports.
2024-01-25 23:42:37 -05:00
05cd00695a
typo fix - calculate_sigmas_scheduler ( #2619 )
...
self.scheduler -> scheduler_name
Co-authored-by: Lt.Dr.Data <lt.dr.data@gmail.com>
2024-01-23 03:47:01 -05:00
4871a36458
Cleanup some unused imports.
2024-01-21 21:51:22 -05:00
78a70fda87
Remove useless import.
2024-01-19 15:38:05 -05:00
d76a04b6ea
Add unfinished ImageOnlyCheckpointSave node to save a SVD checkpoint.
...
This node is unfinished, SVD checkpoints saved with this node will
work with ComfyUI but not with anything else.
2024-01-17 19:46:21 -05:00
f9e55d8463
Only auto enable bf16 VAE on nvidia GPUs that actually support it.
2024-01-15 03:10:22 -05:00
2395ae740a
Make unclip more deterministic.
...
Pass a seed argument note that this might make old unclip images different.
2024-01-14 17:28:31 -05:00
53c8a99e6c
Make server storage the default.
...
Remove --server-storage argument.
2024-01-11 17:21:40 -05:00
977eda19a6
Don't round noise mask.
2024-01-11 03:29:58 -05:00
10f2609fdd
Add InpaintModelConditioning node.
...
This is an alternative to VAE Encode for inpaint that should work with
lower denoise.
This is a different take on #2501
2024-01-11 03:15:27 -05:00
1a57423d30
Fix issue when using multiple t2i adapters with batched images.
2024-01-10 04:00:49 -05:00
6a7bc35db8
Use basic attention implementation for small inputs on old pytorch.
2024-01-09 13:46:52 -05:00
235727fed7
Store user settings/data on the server and multi user support ( #2160 )
...
* wip per user data
* Rename, hide menu
* better error
rework default user
* store pretty
* Add userdata endpoints
Change nodetemplates to userdata
* add multi user message
* make normal arg
* Fix tests
* Ignore user dir
* user tests
* Changed to default to browser storage and add server-storage arg
* fix crash on empty templates
* fix settings added before load
* ignore parse errors
2024-01-08 17:06:44 -05:00
c6951548cf
Update optimized_attention_for_device function for new functions that
...
support masked attention.
2024-01-07 13:52:08 -05:00
aaa9017302
Add attention mask support to sub quad attention.
2024-01-07 04:13:58 -05:00
0c2c9fbdfa
Support attention mask in split attention.
2024-01-06 13:16:48 -05:00
3ad0191bfb
Implement attention mask on xformers.
2024-01-06 04:33:03 -05:00
8c6493578b
Implement noise augmentation for SD 4X upscale model.
2024-01-03 14:27:11 -05:00
ef4f6037cb
Fix model patches not working in custom sampling scheduler nodes.
2024-01-03 12:16:30 -05:00
a7874d1a8b
Add support for the stable diffusion x4 upscaling model.
...
This is an old model.
Load the checkpoint like a regular one and use the new
SD_4XUpscale_Conditioning node.
2024-01-03 03:37:56 -05:00
2c4e92a98b
Fix regression.
2024-01-02 14:41:33 -05:00
5eddfdd80c
Refactor VAE code.
...
Replace constants with downscale_ratio and latent_channels.
2024-01-02 13:24:34 -05:00
a47f609f90
Auto detect out_channels from model.
2024-01-02 01:50:57 -05:00
79f73a4b33
Remove useless code.
2024-01-02 01:50:29 -05:00
1b103e0cb2
Add argument to run the VAE on the CPU.
2023-12-30 05:49:07 -05:00
12e822c6c8
Use function to calculate model size in model patcher.
2023-12-28 21:46:20 -05:00
e1e322cf69
Load weights that can't be lowvramed to target device.
2023-12-28 21:41:10 -05:00
c782144433
Fix clip vision lowvram mode not working.
2023-12-27 13:50:57 -05:00
f21bb41787
Fix taesd VAE in lowvram mode.
2023-12-26 12:52:21 -05:00
61b3f15f8f
Fix lowvram mode not working with unCLIP and Revision code.
2023-12-26 05:02:02 -05:00
d0165d819a
Fix SVD lowvram mode.
2023-12-24 07:13:18 -05:00
a252963f95
--disable-smart-memory now unloads everything like it did originally.
2023-12-23 04:25:06 -05:00
36a7953142
Greatly improve lowvram sampling speed by getting rid of accelerate.
...
Let me know if this breaks anything.
2023-12-22 14:38:45 -05:00
261bcbb0d9
A few missing comfy ops in the VAE.
2023-12-22 04:05:42 -05:00
9a7619b72d
Fix regression with inpaint model.
2023-12-19 02:32:59 -05:00
571ea8cdcc
Fix SAG not working with cfg 1.0
2023-12-18 17:03:32 -05:00
8cf1daa108
Fix SDXL area composition sometimes not using the right pooled output.
2023-12-18 12:54:23 -05:00
2258f85159
Support stable zero 123 model.
...
To use it use the ImageOnlyCheckpointLoader to load the checkpoint and
the new Stable_Zero123 node.
2023-12-18 03:48:04 -05:00
2f9d6a97ec
Add --deterministic option to make pytorch use deterministic algorithms.
2023-12-17 16:59:21 -05:00
e45d920ae3
Don't resize clip vision image when the size is already good.
2023-12-16 03:06:10 -05:00
13e6d5366e
Switch clip vision to manual cast.
...
Make it use the same dtype as the text encoder.
2023-12-16 02:47:26 -05:00
719fa0866f
Set clip vision model in eval mode so it works without inference mode.
2023-12-15 18:53:08 -05:00
574363a8a6
Implement Perp-Neg
2023-12-16 00:28:16 +05:30
a5056cfb1f
Remove useless code.
2023-12-15 01:28:16 -05:00
329c571993
Improve code legibility.
2023-12-14 11:41:49 -05:00
6c5990f7db
Fix cfg being calculated more than once if sampler_cfg_function.
2023-12-13 20:28:04 -05:00
ba04a87d10
Refactor and improve the sag node.
...
Moved all the sag related code to comfy_extras/nodes_sag.py
2023-12-13 16:11:26 -05:00
6761233e9d
Implement Self-Attention Guidance ( #2201 )
...
* First SAG test
* need to put extra options on the model instead of patcher
* no errors and results seem not-broken
* Use @ashen-uncensored formula, which works better!!!
* Fix a crash when using weird resolutions. Remove an unnecessary UNet call
* Improve comments, optimize memory in blur routine
* SAG works with sampler_cfg_function
2023-12-13 15:52:11 -05:00
b454a67bb9
Support segmind vega model.
2023-12-12 19:09:53 -05:00
824e4935f5
Add dtype parameter to VAE object.
2023-12-12 12:03:29 -05:00
32b7e7e769
Add manual cast to controlnet.
2023-12-12 11:32:42 -05:00
3152023fbc
Use inference dtype for unet memory usage estimation.
2023-12-11 23:50:38 -05:00
77755ab8db
Refactor comfy.ops
...
comfy.ops -> comfy.ops.disable_weight_init
This should make it more clear what they actually do.
Some unused code has also been removed.
2023-12-11 23:27:13 -05:00
b0aab1e4ea
Add an option --fp16-unet to force using fp16 for the unet.
2023-12-11 18:36:29 -05:00
ba07cb748e
Use faster manual cast for fp8 in unet.
2023-12-11 18:24:44 -05:00
57926635e8
Switch text encoder to manual cast.
...
Use fp16 text encoder weights for CPU inference to lower memory usage.
2023-12-10 23:00:54 -05:00
340177e6e8
Disable non blocking on mps.
2023-12-10 01:30:35 -05:00
614b7e731f
Implement GLora.
2023-12-09 18:15:26 -05:00
cb63e230b4
Make lora code a bit cleaner.
2023-12-09 14:15:09 -05:00
174eba8e95
Use own clip vision model implementation.
2023-12-09 11:56:31 -05:00
97015b6b38
Cleanup.
2023-12-08 16:02:08 -05:00
a4ec54a40d
Add linear_start and linear_end to model_config.sampling_settings
2023-12-08 02:49:30 -05:00
9ac0b487ac
Make --gpu-only put intermediate values in GPU memory instead of cpu.
2023-12-08 02:35:45 -05:00
efb704c758
Support attention masking in CLIP implementation.
2023-12-07 02:51:02 -05:00
fbdb14d4c4
Cleaner CLIP text encoder implementation.
...
Use a simple CLIP model implementation instead of the one from
transformers.
This will allow some interesting things that would too hackish to implement
using the transformers implementation.
2023-12-06 23:50:03 -05:00
2db86b4676
Slightly faster lora applying.
2023-12-06 05:13:14 -05:00
1bbd65ab30
Missed this one.
2023-12-05 12:48:41 -05:00
9b655d4fd7
Fix memory issue with control loras.
2023-12-04 21:55:19 -05:00
26b1c0a771
Fix control lora on fp8.
2023-12-04 13:47:41 -05:00
be3468ddd5
Less useless downcasting.
2023-12-04 12:53:46 -05:00
ca82ade765
Use .itemsize to get dtype size for fp8.
2023-12-04 11:52:06 -05:00
31b0f6f3d8
UNET weights can now be stored in fp8.
...
--fp8_e4m3fn-unet and --fp8_e5m2-unet are the two different formats
supported by pytorch.
2023-12-04 11:10:00 -05:00
af365e4dd1
All the unet ops with weights are now handled by comfy.ops
2023-12-04 03:12:18 -05:00
61a123a1e0
A different way of handling multiple images passed to SVD.
...
Previously when a list of 3 images [0, 1, 2] was used for a 6 frame video
they were concated like this:
[0, 1, 2, 0, 1, 2]
now they are concated like this:
[0, 0, 1, 1, 2, 2]
2023-12-03 03:31:47 -05:00
c97be4db91
Support SD2.1 turbo checkpoint.
2023-11-30 19:27:03 -05:00
983ebc5792
Use smart model management for VAE to decrease latency.
2023-11-28 04:58:51 -05:00
c45d1b9b67
Add a function to load a unet from a state dict.
2023-11-27 17:41:29 -05:00
f30b992b18
.sigma and .timestep now return tensors on the same device as the input.
2023-11-27 16:41:33 -05:00
13fdee6abf
Try to free memory for both cond+uncond before inference.
2023-11-27 14:55:40 -05:00
be71bb5e13
Tweak memory inference calculations a bit.
2023-11-27 14:04:16 -05:00
39e75862b2
Fix regression from last commit.
2023-11-26 03:43:02 -05:00
50dc39d6ec
Clean up the extra_options dict for the transformer patches.
...
Now everything in transformer_options gets put in extra_options.
2023-11-26 03:13:56 -05:00
5d6dfce548
Fix importing diffusers unets.
2023-11-24 20:35:29 -05:00
3e5ea74ad3
Make buggy xformers fall back on pytorch attention.
2023-11-24 03:55:35 -05:00
871cc20e13
Support SVD img2vid model.
2023-11-23 19:41:33 -05:00
410bf07771
Make VAE memory estimation take dtype into account.
2023-11-22 18:17:19 -05:00
32447f0c39
Add sampling_settings so models can specify specific sampling settings.
2023-11-22 17:24:00 -05:00
c3ae99a749
Allow controlling downscale and upscale methods in PatchModelAddDownscale.
2023-11-22 03:23:16 -05:00
72741105a6
Remove useless code.
2023-11-21 17:27:28 -05:00
6a491ebe27
Allow model config to preprocess the vae state dict on load.
2023-11-21 16:29:18 -05:00
cd4fc77d5f
Add taesd and taesdxl to VAELoader node.
...
They will show up if both the taesd_encoder and taesd_decoder or taesdxl
model files are present in the models/vae_approx directory.
2023-11-21 12:54:19 -05:00
ce67dcbcda
Make it easy for models to process the unet state dict on load.
2023-11-20 23:17:53 -05:00
d9d8702d8d
percent_to_sigma now returns a float instead of a tensor.
2023-11-18 23:20:29 -05:00
0cf4e86939
Add some command line arguments to store text encoder weights in fp8.
...
Pytorch supports two variants of fp8:
--fp8_e4m3fn-text-enc (the one that seems to give better results)
--fp8_e5m2-text-enc
2023-11-17 02:56:59 -05:00
107e78b1cb
Add support for loading SSD1B diffusers unet version.
...
Improve diffusers model detection.
2023-11-16 23:12:55 -05:00
7e3fe3ad28
Make deep shrink behave like it should.
2023-11-16 15:26:28 -05:00
9f00a18095
Fix potential issues.
2023-11-16 14:59:54 -05:00
7ea6bb038c
Print warning when controlnet can't be applied instead of crashing.
2023-11-16 12:57:12 -05:00
dcec1047e6
Invert the start and end percentages in the code.
...
This doesn't affect how percentages behave in the frontend but breaks
things if you relied on them in the backend.
percent_to_sigma goes from 0 to 1.0 instead of 1.0 to 0 for less confusion.
Make percent 0 return an extremely large sigma and percent 1.0 return a
zero one to fix imprecision.
2023-11-16 04:23:44 -05:00
57eea0efbb
heunpp2 sampler.
2023-11-14 23:50:55 -05:00
728613bb3e
Fix last pr.
2023-11-14 14:41:31 -05:00
ec3d0ab432
Merge branch 'master' of https://github.com/Jannchie/ComfyUI
2023-11-14 14:38:07 -05:00
c962884a5c
Make bislerp work on GPU.
2023-11-14 11:38:36 -05:00
420beeeb05
Clean up and refactor sampler code.
...
This should make it much easier to write custom nodes with kdiffusion type
samplers.
2023-11-14 00:39:34 -05:00
f2e49b1d57
fix: adaptation to older versions of pytroch
2023-11-14 14:32:05 +09:00
94cc718e9c
Add a way to add patches to the input block.
2023-11-14 00:08:12 -05:00
7339479b10
Disable xformers when it can't load properly.
2023-11-13 12:31:10 -05:00
4781819a85
Make memory estimation aware of model dtype.
2023-11-12 04:28:26 -05:00
dd4ba68b6e
Allow different models to estimate memory usage differently.
2023-11-12 04:03:52 -05:00
2c9dba8dc0
sampling_function now has the model object as the argument.
2023-11-12 03:45:10 -05:00
8d80584f6a
Remove useless argument from uni_pc sampler.
2023-11-12 01:25:33 -05:00
248aa3e563
Fix bug.
2023-11-11 12:20:16 -05:00
4a8a839b40
Add option to use in place weight updating in ModelPatcher.
2023-11-11 01:11:12 -05:00
412d3ff57d
Refactor.
2023-11-11 01:11:06 -05:00
58d5d71a93
Working RescaleCFG node.
...
This was broken because of recent changes so I fixed it and moved it from
the experiments repo.
2023-11-10 20:52:10 -05:00
3e0033ef30
Fix model merge bug.
...
Unload models before getting weights for model patching.
2023-11-10 03:19:05 -05:00
002aefa382
Support lcm models.
...
Use the "lcm" sampler to sample them, you also have to use the
ModelSamplingDiscrete node to set them as lcm models to use them properly.
2023-11-09 18:30:22 -05:00
ec12000136
Add support for full diff lora keys.
2023-11-08 22:05:31 -05:00
064d7583eb
Add a CONDConstant for passing non tensor conds to unet.
2023-11-08 01:59:09 -05:00
794dd2064d
Fix typo.
2023-11-07 23:41:55 -05:00
0a6fd49a3e
Print leftover keys when using the UNETLoader.
2023-11-07 22:15:55 -05:00
fe40109b57
Fix issue with object patches not being copied with patcher.
2023-11-07 22:15:15 -05:00
a527d0c795
Code refactor.
2023-11-07 19:33:40 -05:00
2a23ba0b8c
Fix unet ops not entirely on GPU.
2023-11-07 04:30:37 -05:00
844dbf97a7
Add: advanced->model->ModelSamplingDiscrete node.
...
This allows changing the sampling parameters of the model (eps or vpred)
or set the model to use zsnr.
2023-11-07 03:28:53 -05:00
656c0b5d90
CLIP code refactor and improvements.
...
More generic clip model class that can be used on more types of text
encoders.
Don't apply weighting algorithm when weight is 1.0
Don't compute an empty token output when it's not needed.
2023-11-06 14:17:41 -05:00
b3fcd64c6c
Make SDTokenizer class work with more types of tokenizers.
2023-11-06 01:09:18 -05:00
7e455adc07
fix unet_wrapper_function name in ModelPatcher
2023-11-05 17:11:44 +08:00
1ffa8858e7
Move model sampling code to comfy/model_sampling.py
2023-11-04 01:32:23 -04:00
ae2acfc21b
Don't convert Nan to zero.
...
Converting Nan to zero is a bad idea because it makes it hard to tell when
something went wrong.
2023-11-03 13:13:15 -04:00
d2e27b48f1
sampler_cfg_function now gets the noisy output as argument again.
...
This should make things that use sampler_cfg_function behave like before.
Added an input argument for those that want the denoised output.
This means you can calculate the x0 prediction of the model by doing:
(input - cond) for example.
2023-11-01 21:24:08 -04:00
2455aaed8a
Allow model or clip to be None in load_lora_for_models.
2023-11-01 20:27:20 -04:00
ecb80abb58
Allow ModelSamplingDiscrete to be instantiated without a model config.
2023-11-01 19:13:03 -04:00
e73ec8c4da
Not used anymore.
2023-11-01 00:01:30 -04:00
111f1b5255
Fix some issues with sampling precision.
2023-10-31 23:49:29 -04:00
7c0f255de1
Clean up percent start/end and make controlnets work with sigmas.
2023-10-31 22:14:32 -04:00
a268a574fa
Remove a bunch of useless code.
...
DDIM is the same as euler with a small difference in the inpaint code.
DDIM uses randn_like but I set a fixed seed instead.
I'm keeping it in because I'm sure if I remove it people are going to
complain.
2023-10-31 18:11:29 -04:00
1777b54d02
Sampling code changes.
...
apply_model in model_base now returns the denoised output.
This means that sampling_function now computes things on the denoised
output instead of the model output. This should make things more consistent
across current and future models.
2023-10-31 17:33:43 -04:00
c837a173fa
Fix some memory issues in sub quad attention.
2023-10-30 15:30:49 -04:00
125b03eead
Fix some OOM issues with split attention.
2023-10-30 13:14:11 -04:00
a12cc05323
Add --max-upload-size argument, the default is 100MB.
2023-10-29 03:55:46 -04:00
2a134bfab9
Fix checkpoint loader with config.
2023-10-27 22:13:55 -04:00
e60ca6929a
SD1 and SD2 clip and tokenizer code is now more similar to the SDXL one.
2023-10-27 15:54:04 -04:00
6ec3f12c6e
Support SSD1B model and make it easier to support asymmetric unets.
2023-10-27 14:45:15 -04:00
434ce25ec0
Restrict loading embeddings from embedding folders.
2023-10-27 02:54:13 -04:00
723847f6b3
Faster clip image processing.
2023-10-26 01:53:01 -04:00
a373367b0c
Fix some OOM issues with split and sub quad attention.
2023-10-25 20:17:28 -04:00
7fbb217d3a
Fix uni_pc returning noisy image when steps <= 3
2023-10-25 16:08:30 -04:00
3783cb8bfd
change 'c_adm' to 'y' in ControlNet.get_control
2023-10-25 08:24:32 -05:00
d1d2fea806
Pass extra conds directly to unet.
2023-10-25 00:07:53 -04:00
036f88c621
Refactor to make it easier to add custom conds to models.
2023-10-24 23:31:12 -04:00
3fce8881ca
Sampling code refactor to make it easier to add more conds.
2023-10-24 03:38:41 -04:00
8594c8be4d
Empty the cache when torch cache is more than 25% free mem.
2023-10-22 13:58:12 -04:00
8b65f5de54
attention_basic now works with hypertile.
2023-10-22 03:59:53 -04:00
e6bc42df46
Make sub_quad and split work with hypertile.
2023-10-22 03:51:29 -04:00
a0690f9df9
Fix t2i adapter issue.
2023-10-21 20:31:24 -04:00
9906e3efe3
Make xformers work with hypertile.
2023-10-21 13:23:03 -04:00
4185324a1d
Fix uni_pc sampler math. This changes the images this sampler produces.
2023-10-20 04:16:53 -04:00
e6962120c6
Make sure cond_concat is on the right device.
2023-10-19 01:14:25 -04:00
45c972aba8
Refactor cond_concat into conditioning.
2023-10-18 20:36:58 -04:00
430a8334c5
Fix some potential issues.
2023-10-18 19:48:36 -04:00
782a24fce6
Refactor cond_concat into model object.
2023-10-18 16:48:37 -04:00
0d45a565da
Fix memory issue related to control loras.
...
The cleanup function was not getting called.
2023-10-18 02:43:01 -04:00
d44a2de49f
Make VAE code closer to sgm.
2023-10-17 15:18:51 -04:00
23680a9155
Refactor the attention stuff in the VAE.
2023-10-17 03:19:29 -04:00
c8013f73e5
Add some Quadro cards to the list of cards with broken fp16.
2023-10-16 16:48:46 -04:00
bb064c9796
Add a separate optimized_attention_masked function.
2023-10-16 02:31:24 -04:00
fd4c5f07e7
Add a --bf16-unet to test running the unet in bf16.
2023-10-13 14:51:10 -04:00
9a55dadb4c
Refactor code so model can be a dtype other than fp32 or fp16.
2023-10-13 14:41:17 -04:00
88733c997f
pytorch_attention_enabled can now return True when xformers is enabled.
2023-10-11 21:30:57 -04:00
20d3852aa1
Pull some small changes from the other repo.
2023-10-11 20:38:48 -04:00
ac7d8cfa87
Allow attn_mask in attention_pytorch.
2023-10-11 20:38:48 -04:00
1a4bd9e9a6
Refactor the attention functions.
...
There's no reason for the whole CrossAttention object to be repeated when
only the operation in the middle changes.
2023-10-11 20:38:48 -04:00
8cc75c64ff
Let unet wrapper functions have .to attributes.
2023-10-11 01:34:38 -04:00
5e885bd9c8
Cleanup.
2023-10-10 21:46:53 -04:00
851bb87ca9
Merge branch 'taesd_safetensors' of https://github.com/mochiya98/ComfyUI
2023-10-10 21:42:35 -04:00
9eb621c95a
Supports TAESD models in safetensors format
2023-10-10 13:21:44 +09:00
d1a0abd40b
Merge branch 'input-directory' of https://github.com/jn-jairo/ComfyUI
2023-10-09 01:53:29 -04:00
72188dffc3
load_checkpoint_guess_config can now optionally output the model.
2023-10-06 13:48:18 -04:00
63e5fd1790
Option to input directory
2023-10-04 19:45:15 -03:00
9bfec2bdbf
Fix quality loss due to low precision
2023-10-04 15:40:59 +02:00
0f17993d05
fix: typo in extra sampler
2023-09-29 06:09:59 +03:00
66756de100
Add SamplerDPMPP_2M_SDE node.
2023-09-28 21:56:23 -04:00
71713888c4
Print missing VAE keys.
2023-09-28 00:54:57 -04:00
d234ca558a
Add missing samplers to KSamplerSelect.
2023-09-28 00:17:03 -04:00
1adcc4c3a2
Add a SamplerCustom Node.
...
This node takes a list of sigmas and a sampler object as input.
This lets people easily implement custom schedulers and samplers as nodes.
More nodes will be added to it in the future.
2023-09-27 22:21:18 -04:00
bf3fc2f1b7
Refactor sampling related code.
2023-09-27 16:45:22 -04:00
fff491b032
Model patches can now know which batch is positive and negative.
2023-09-27 12:04:07 -04:00
1d6dd83184
Scheduler code refactor.
2023-09-26 17:07:07 -04:00
446caf711c
Sampling code refactor.
2023-09-26 13:45:15 -04:00
76cdc809bf
Support more controlnet models.
2023-09-23 18:47:46 -04:00
ae87543653
Merge branch 'cast_intel' of https://github.com/simonlui/ComfyUI
2023-09-23 00:57:17 -04:00
eec449ca8e
Allow Intel GPUs to LoRA cast on GPU since it supports BF16 natively.
2023-09-22 21:11:27 -07:00
afa2399f79
Add a way to set output block patches to modify the h and hsp.
2023-09-22 20:26:47 -04:00
492db2de8d
Allow having a different pooled output for each image in a batch.
2023-09-21 01:14:42 -04:00
1cdfb3dba4
Only do the cast on the device if the device supports it.
2023-09-20 17:52:41 -04:00
7c9a92f552
Don't depend on torchvision.
2023-09-19 13:12:47 -04:00
2b6b178173
Added support for lanczos scaling
2023-09-19 10:40:38 +02:00
b92bf8196e
Do lora cast on GPU instead of CPU for higher performance.
2023-09-18 23:04:49 -04:00
321c5fa295
Enable pytorch attention by default on xpu.
2023-09-17 04:09:19 -04:00
61b1f67734
Support models without previews.
2023-09-16 12:59:54 -04:00
43d4935a1d
Add cond_or_uncond array to transformer_options so hooks can check what is
...
cond and what is uncond.
2023-09-15 22:21:14 -04:00
415abb275f
Add DDPM sampler.
2023-09-15 19:22:47 -04:00
94e4fe39d8
This isn't used anywhere.
2023-09-15 12:03:03 -04:00
44361f6344
Support for text encoder models that need attention_mask.
2023-09-15 02:02:05 -04:00
0d8f376446
Set last layer on SD2.x models uses the proper indexes now.
...
Before I had made the last layer the penultimate layer because some
checkpoints don't have them but it's not consistent with the others models.
TLDR: for SD2.x models only: CLIPSetLastLayer -1 is now -2.
2023-09-14 20:28:22 -04:00
0966d3ce82
Don't run text encoders on xpu because there are issues.
2023-09-14 12:16:07 -04:00
3039b08eb1
Only parse command line args when main.py is called.
2023-09-13 11:38:20 -04:00
ed58730658
Don't leave very large hidden states in the clip vision output.
2023-09-12 15:09:10 -04:00
fb3b728203
Fix issue where autocast fp32 CLIP gave different results from regular.
2023-09-11 21:49:56 -04:00
7d401ed1d0
Add ldm format support to UNETLoader.
2023-09-11 16:36:50 -04:00
e85be36bd2
Add a penultimate_hidden_states to the clip vision output.
2023-09-08 14:06:58 -04:00
1e6b67101c
Support diffusers format t2i adapters.
2023-09-08 11:36:51 -04:00
326577d04c
Allow cancelling of everything with a progress bar.
2023-09-07 23:37:03 -04:00
f88f7f413a
Add a ConditioningSetAreaPercentage node.
2023-09-06 03:28:27 -04:00
1938f5c5fe
Add a force argument to soft_empty_cache to force a cache empty.
2023-09-04 00:58:18 -04:00
7746bdf7b0
Merge branch 'generalize_fixes' of https://github.com/simonlui/ComfyUI
2023-09-04 00:43:11 -04:00
comfyanonymous
Chenlei Hu
Silver
BiologicalExplosion
maedtb
bymyself
Jukka Seppänen
zhoufan2956
HishamC
Pam
catboxanon
Dr.Lt.Data
KarryCharon
filtered
chaObserv
Sergii Dymchenko
Jedrzej Kosinski
Andrew Kvochko
blepping
Huazhong Ji
Simon Lui
zhangp365
City
Raphael Walker
Haoming
Michael Kupchick
comfyanonymous
contentis
Aarni Koskela
PsychoLogicAu
svdc
Kadir Nar
Jonathan Avila
yoinked
pharmapsychotic
Robin Huang
Jim Winkens
Scorpinaus
David
Svein Ove Aas
Ashen
Xrvk
Matthew Turnshek
guill
ljleb
Jaret Burkett
TTPlanetPig
来新璐
a-One-Fan
Alexander Brown
喵哩个咪
Thomas Ward
Extraltodeus
Alex "mcmonkey" Goodwin
Mario Klingemann
Janek Mann
Dango233
JettHu
Garrett Sutula
kk-89
Shiimizu
logtd
blepping
pythongosssss
Hari
Jianqi Pan
gameltb
Yukimasa Funaoka
Jairo Correa
badayvedat
MoonRide303