ca72eb1515
Add 1328 native resolution for Qwen Image training ( #749 )
...
* Add 1328 native resolution for Qwen Image training
Qwen-Image and Qwen-Image-2512 have a native 1:1 resolution of 1328x1328
as documented in the official model card's aspect ratio table. Adding it
to the resolution buckets and UI allows training at the model's native
resolution for improved quality.
* Revert example config change (24GB OOM at 1328)
2026-03-28 12:09:15 -06:00
5ce87fa48b
Version bump
2026-03-27 20:26:31 -06:00
740657e25e
Improve dataset uploader. Upload the files one at a time instead of one huge chunk. Show progress for each file.
2026-03-27 09:26:22 -06:00
f85bf065bf
Use pooler embeddings for DFE v6 with dino v3
2026-03-27 07:02:07 -06:00
a802014ec5
Update the torch versions in the README
2026-03-26 12:15:32 -06:00
2782df02c3
Allow HF_HUB_ENABLE_HF_TRANSFER to be set via env variable
2026-03-26 10:45:49 -06:00
2c8d2acdcb
On jobs table, sort idle jobs by last updated so recent active ones are at the top
2026-03-26 10:33:17 -06:00
9a77389653
On a new training job, or when editing one, load everything before allowing editing
2026-03-26 10:23:42 -06:00
a7bb4ddb2c
Work on loss graph. Add smoothed overlay. Allow user to hilite a secton of the graph to zoom into.
2026-03-26 10:05:09 -06:00
401f7df425
Merge branch 'main' of github.com:ostris/ai-toolkit
2026-03-26 09:11:50 -06:00
4df3b0463f
Save job pid to the database and sing sigint to kill it when stopping so it stops immediatly.
2026-03-26 09:10:37 -06:00
489b194231
Fix CPU/CUDA device mismatch in Klein edit control image encoding ( #742 )
...
When training Klein models with a `control_path` (edit/kontext-style
paired datasets), `encode_image_refs()` returns tensors that reside on
the VAE's device (CPU, since the VAE weights are loaded via
`load_file(..., device="cpu")` and are never explicitly moved to the
training device). Concatenating those CPU tensors with the training
latents (`packed_latents`) that live on CUDA raises:
RuntimeError: Expected all tensors to be on the same device
Fix: move `img_cond_seq` and `img_cond_seq_ids` to the same device
(and dtype) as `img_input` / `img_input_ids` before concatenation.
Co-authored-by: HuangYuChuh <HuangYuChuh@users.noreply.github.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-25 11:45:38 -06:00
89d2090962
Fixed race condition that would occasionally set the dataset path to the first one when editing a job
2026-03-25 11:22:42 -06:00
3f7a3d8d87
Shorten stal action to 3 months
2026-03-25 10:42:18 -06:00
45647c15d3
Added github actions to close stale issues automatically. Hopefully it doesnt break things
2026-03-25 10:25:56 -06:00
899ee528f9
Update git ignore
2026-03-25 10:16:36 -06:00
5d5a8ef9da
Fixed issue with deleting datasets and jobs with newer version of node.js. Bumped minimum version of node js to 20
2026-03-25 10:04:36 -06:00
dfde30f231
Fix issue with ltx2 custom te repo path
2026-03-25 09:50:18 -06:00
b8000dbcbc
Bump version
2026-03-25 08:18:42 -06:00
54f4732c9b
Fix the bug in temporal_compression data loader ( #754 )
2026-03-25 08:16:44 -06:00
7f3309b291
Add support for audo frame count so datasets can have varrying length videos. Varous ltx 2.3 VAE optimizations such as removing tiling articacts, and doing frame split encoding to reduce vram on encoding/decoding.
2026-03-24 12:20:09 -06:00
4ad14d211a
Add an import config button ( #733 )
2026-03-23 15:41:27 -06:00
7a0bbca5b1
Fix random_noise_multiplier ( #738 )
...
Apply random_noise_multiplier to noise.
2026-03-23 15:22:16 -06:00
99a4a5887b
Fix Qwen attention mask crash with diffusers >=0.37 ( #748 )
...
* Fix Qwen Image mask handling
* Fix Qwen attention mask crash with diffusers >=0.37
diffusers v0.37 (PR #12987 ) optimizes all-ones attention masks to None
in encode_prompt() when there is no padding. This breaks ai-toolkit's
Qwen extensions which call .to() on the mask unconditionally.
Fix: reconstruct the all-ones mask at the boundary (get_prompt_embeds)
right after encode_prompt() returns. This keeps the rest of the code
unchanged and works with both old and new diffusers versions.
Also removes redundant duplicate mask assignments in qwen_image_edit
and qwen_image_edit_plus.
Fixes #740
2026-03-23 14:43:08 -06:00
295094b4b5
Fixed new breaking change in diffusers with with qwen image
2026-03-23 14:10:55 -06:00
5642b656b9
Fix audio issues with ltx2 models. Silent codec fails now raised. Auto convert surround sound audio to stereo. Invalidate old caches just to be safe so they recache now.
2026-03-23 20:08:33 +00:00
561e6f201c
Fixed an issue with ltx 2.3 i2v training
2026-03-23 12:41:18 -06:00
330059d8a1
version bump
2026-03-23 11:01:16 -06:00
e91827f9be
Change gemma repo to lightricks one that is not gated
2026-03-23 11:00:32 -06:00
253cb31362
Fix issue with video and images with no audio on ltx models
2026-03-22 22:09:23 -06:00
4a3d317e2b
Fix issue with using the default text encoder with ltx 2.3
2026-03-22 18:53:59 -06:00
859635e95b
Add support for training LTX 2.3 ( #745 )
...
* Initial support for ltx 2.3. Still needs a lot of testing to make sure it is all right.
* bump version
* Handle lora renaming keys for new ltx 2.3 layers
2026-03-22 17:56:59 -06:00
7e1fdc3844
Remove the 0.1 floor for amplification
2026-03-22 09:01:58 -06:00
0f075fc45e
Adjust signal amplification target. Allow signal amplification strength in config.
2026-03-22 08:30:13 -06:00
dcd98dc0d5
Add signal amplification
2026-03-21 07:44:18 -06:00
35b1cde3cb
Fixed issue on z-image that prevented training at a larger batch size
2026-03-10 15:43:25 -06:00
4909b809c7
Fixed issue with audio loss multiplier.
2026-03-10 15:16:09 -06:00
06ef3d343a
add ability to use batch noise correction during training
2026-03-10 09:05:57 -06:00
b04c64e0f8
Add a dino version of DFE
2026-03-04 08:20:37 -07:00
9dee42fc09
Updated supporters
2026-03-03 08:04:37 -07:00
35978df8a3
Adjust defaults for ui graph to get and show all losses
2026-03-02 10:27:15 -07:00
57d407cfd4
Add support for training lodestones/Zeta-Chroma
2026-03-01 12:52:29 -07:00
40f995f616
Add method to do continuious lora merging in for low vram full finetuning.
2026-02-26 09:00:41 -07:00
de7d22c9be
Version bump
2026-02-19 11:58:15 -07:00
1c74ca5d22
Add audio_loss_multiplier to scale audio loss to larger values if desired.
2026-02-19 11:57:44 -07:00
3632656cda
make DFE work with more VAEs
2026-02-18 09:46:37 -07:00
a055947d56
Add signal_correction_noise_scale to config to scale the signal correction strength
2026-02-07 12:04:21 -07:00
454722cc97
Add signal correction noise
2026-02-07 09:49:55 -07:00
e82cf6eec2
Fixed issue that prevented full fine-tuning of flux2 models when using gradient checkpointing
2026-02-06 16:18:43 -07:00
1422789452
Improved the method to augment random noise
2026-02-06 15:44:10 -07:00
Rayane
Jaret Burkett
科林 KELIN
Rodrigo Reis
Rasmus Lerdorf
Remix