349098a7cc
Add PixlStash remote dataset source plugin
...
A reference plugin for the remote dataset source framework that browses and
imports training images from a PixlStash server.
- Browse characters and picture sets, import selected groups as a dataset.
- Captions are built from each picture's natural-language description and WD14
tags, read per-picture from GET /pictures/{id}/metadata (the listing
endpoints only return scalar grid fields).
- Supports SSL and configurable API paths.
Also un-ignores extensions/pixlstash in .gitignore so the plugin is tracked.
2026-08-07 12:30:08 +00:00
418fc82bdc
Add support for remote dataset source plugins
...
Introduce a plugin system that lets datasets be browsed and imported from
external services directly in the UI.
- toolkit/dataset_sources/: base class, registry, and a CLI bridge that the
Next.js API routes call to list plugins, browse items, fetch thumbnails,
and build import job configs. Plugins live in extensions/ and self-register
from their __init__.py, so they are only loaded on demand.
- ui: a RemoteSourceBrowseModal plus /api/datasets/remote/* routes for
listing plugins, browsing, thumbnails, and importing, with the settings
page driven by each plugin's settings_schema.
The framework is plugin-agnostic; no plugin is bundled here.
2026-08-07 12:05:32 +00:00
685ce37a8d
Adjust sample timestep sigmas to be model evals for h3 for 1 extra step
2026-08-06 19:54:04 -06:00
9171d5ec1d
Update training adapter path
2026-08-06 11:21:28 -06:00
71625d1207
Add the alpha version of the MiniMax H3 trianing adapter and set it as the new default training method.
2026-08-06 11:11:31 -06:00
b904b99705
varsion bump
2026-08-06 09:25:35 -06:00
b811636ae4
Check for finite vs isnan on loss before backpropigating. to catch infinity overflows
2026-08-06 09:18:56 -06:00
edacd406b3
Handle minimax loading of non pruned model
2026-08-06 09:18:00 -06:00
7309db4d74
Fix layer offloaded adaln projection layer move
2026-08-06 08:10:20 -06:00
139a38f5bd
Upcast adaln pruned layers to fp32 on h3 to prevent overflow.
2026-08-06 07:51:15 -06:00
1e1418b22c
Apply sigma to contrastive guidance to balance loss better. Prevent noise grads on images/non audio datasets. Prep for training adapters on MiniMax H3
2026-08-05 13:27:12 -06:00
9065951da3
Version bump
2026-08-04 22:14:54 -06:00
3afa270ab5
Fix audio losses for DOP and other preservation losses
2026-08-04 22:08:48 -06:00
0f9094db95
Fix audio loss when doing do_guidance_loss
2026-08-04 21:43:31 -06:00
d870e9b68a
Fix backwards compatability for older versions of triton
2026-08-04 21:07:57 -06:00
a8d67ecd90
Version Bump
2026-08-04 16:28:04 -06:00
9fc1f208df
Remove adaln_proj from the network modules for minimax_h3
2026-08-04 16:19:46 -06:00
183433ae8e
Make MiniMax H3 default to using contrastive guidance loss to prevent distillation breakdown
2026-08-04 16:13:45 -06:00
00a93e3830
Add dataset flag to cache the raw tensors
2026-08-04 15:32:48 -06:00
8a0bcf1ffe
Fix issue loading H3 test encoder on older GPUS
2026-08-04 15:30:29 -06:00
dc29ae1187
Version Bump
2026-08-04 09:27:46 -06:00
d20a17c10e
Limit max tokens to 512. Allow override with model kwargs.
2026-08-04 07:55:21 -06:00
602306da77
Add gradient checkpointing to vae
2026-08-04 07:54:51 -06:00
18f5810d6c
Adjust default alpha for h3
2026-08-03 20:15:28 -06:00
a9a04547e9
Dont move encoders on and off device when caching until the first instance of needing to process a cache item.
2026-08-03 15:54:30 -06:00
41676bb258
Queue up videos with multiple threads when caching latents so the VAE is not waiting on videos to process
2026-08-03 15:41:36 -06:00
546eb7daff
Look for existing models in folders recursivly
2026-08-03 15:39:55 -06:00
d3a3f70a2a
Speed up quantization processing on H3
2026-08-03 15:21:39 -06:00
88ac27fc8f
Handle images with MiniMax H3.
2026-08-03 11:58:44 -06:00
bf739ff966
Fix issue with layer offloading with MinMax H3
2026-08-03 11:32:36 -06:00
9d614a51fb
Hard fail if a step in the docker build fails.
2026-08-03 10:38:31 -06:00
8502a845b1
Add support for MiniMax H3 T2V and I2V training
2026-08-03 10:17:39 -06:00
73cab2acf5
Reworked merge in out of loras with convrot weights for better roundtrip accuracy.
2026-08-02 08:29:17 -06:00
a6f6b6b896
Fixed experiment opt name match
2026-08-01 17:43:36 -06:00
6b95282097
Recover from issue when a video model first fram may not have been cached properly
2026-08-01 12:00:33 -06:00
5baa495585
DFE optimizations
2026-08-01 10:06:16 -06:00
fc78b07332
Attach the ema to the base model so it can be used on specific models
2026-08-01 10:05:42 -06:00
c68e58083f
Add code for automagic experiment
2026-08-01 10:04:46 -06:00
497014bf5d
Set i2v to default to false when omitted from the config
2026-07-31 19:11:58 -06:00
038f24e8c3
Revert torchao back to older version
2026-07-31 16:34:06 -06:00
6e7bc81241
Allow random noise shift with video latents.
2026-07-31 05:58:08 -06:00
ddc69745fe
Update docker build image with newer dependencies
2026-07-30 14:20:31 -06:00
2cab330392
Merge pull request #986 from ostris/dev
...
Add AI Toolkit Manager script that auto installs/runs/and updates AI Toolkit.
2026-07-30 11:55:13 -06:00
c8636478f9
Version Bump
2026-07-30 11:48:01 -06:00
7b2386c096
Handle video codecs that fail in opencv
2026-07-29 21:01:57 -06:00
9021caa723
Merge branch 'dev' of github.com:ostris/ai-toolkit into dev
2026-07-29 17:50:50 -06:00
3f8afcac7e
Allow dataloader to encode first frame with the text embeddings if the model needs it.
2026-07-29 17:50:45 -06:00
23f1ebfb76
Update to high speed xet env var
2026-07-29 17:47:53 -06:00
3d472de2f1
Reworked spawning on dataloader so windows and mac can use multiple data loader workers now
2026-07-29 17:47:04 -06:00
65443cfffa
Add information about the new manager to the README
2026-07-29 10:29:36 -06:00
Gaute Lindkvist
Jaret Burkett