Jaret Burkett
89102f76dc
Allow user to pick if reference images are sent in the image or video reference stream for Hidream H3 ref2va
2026-08-19 13:32:06 -06:00
Jaret Burkett
afd1d92722
Fix issue with double sampling qwen vl video frames for hidream h3 for reference videos
2026-08-17 15:39:07 -06:00
Jaret Burkett
42dfe9c661
Improve video frame reader
2026-08-16 10:21:11 -06:00
Jaret Burkett
2cbc2bb097
On minimax h3, trim cached text embeddings to max tokens if they are longer than specified
2026-08-16 08:42:01 -06:00
Jaret Burkett
151ad0e959
Adjust MiniMah h3 sizing for videos to downscale to match
2026-08-15 07:50:42 -06:00
Jaret Burkett
127d6f626d
Rework img/video reference in Minimax H3 to more closely match the comfy ui implementation.
2026-08-15 07:13:54 -06:00
Jaret Burkett
97bf49edad
Add support for video references in MiniMax H3 ref2va
2026-08-15 06:18:09 -06:00
Jaret Burkett
6ea281973d
Add support for MiniMax H3 Ref2Vid training
2026-08-13 12:36:03 -06:00
Jaret Burkett
a1ddeeef13
Fixed issue with offloading text encoder on ltx 2.5
2026-08-12 10:46:58 -06:00
Jaret Burkett
cbf910ac02
Add support for LTX 2.5
2026-08-12 05:50:30 -06:00
Jaret Burkett
924c426675
Resolve the new ltx 2.5 path subfolder for ltx 2.5. Sill checking compatability.
2026-08-11 13:30:17 -06:00
Jaret Burkett
685ce37a8d
Adjust sample timestep sigmas to be model evals for h3 for 1 extra step
2026-08-06 19:54:04 -06:00
Jaret Burkett
edacd406b3
Handle minimax loading of non pruned model
2026-08-06 09:18:00 -06:00
Jaret Burkett
7309db4d74
Fix layer offloaded adaln projection layer move
2026-08-06 08:10:20 -06:00
Jaret Burkett
139a38f5bd
Upcast adaln pruned layers to fp32 on h3 to prevent overflow.
2026-08-06 07:51:15 -06:00
Jaret Burkett
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
Jaret Burkett
3afa270ab5
Fix audio losses for DOP and other preservation losses
2026-08-04 22:08:48 -06:00
Jaret Burkett
0f9094db95
Fix audio loss when doing do_guidance_loss
2026-08-04 21:43:31 -06:00
Jaret Burkett
d20a17c10e
Limit max tokens to 512. Allow override with model kwargs.
2026-08-04 07:55:21 -06:00
Jaret Burkett
602306da77
Add gradient checkpointing to vae
2026-08-04 07:54:51 -06:00
Jaret Burkett
546eb7daff
Look for existing models in folders recursivly
2026-08-03 15:39:55 -06:00
Jaret Burkett
d3a3f70a2a
Speed up quantization processing on H3
2026-08-03 15:21:39 -06:00
Jaret Burkett
88ac27fc8f
Handle images with MiniMax H3.
2026-08-03 11:58:44 -06:00
Jaret Burkett
bf739ff966
Fix issue with layer offloading with MinMax H3
2026-08-03 11:32:36 -06:00
Jaret Burkett
8502a845b1
Add support for MiniMax H3 T2V and I2V training
2026-08-03 10:17:39 -06:00
Jaret Burkett
92bdb6e473
Add support for Mage-Flow and Mage-Flow Edit
2026-07-25 11:34:20 -06:00
Jaret Burkett
f1bc6508ad
Fix issue with qwen image edit models.
2026-07-17 09:14:04 -06:00
Jaret Burkett
5cb54ba9cc
Allow setting weight saving flag on hidream_o1
2026-07-16 07:40:25 -06:00
DasPauluteli
a92f18bf71
krea2: don't hardcode the NVIDIA-only cuDNN SDPA backend ( #933 )
...
* krea2: don't hardcode NVIDIA-only cuDNN SDPA backend
The krea2 attention() forced SDPBackend.CUDNN_ATTENTION, which is
NVIDIA-only. On non-NVIDIA backends (AMD ROCm, Intel XPU, Apple MPS)
every forward pass fails with 'RuntimeError: No available kernel.
Aborting execution.', so Krea 2 LoRA training cannot run at all there.
Pass a priority list [CUDNN, FLASH, EFFICIENT, MATH] instead. NVIDIA
still selects cuDNN; other backends fall back to flash/efficient/math.
Verified training end-to-end on an AMD Radeon 8060S (gfx1151, ROCm 7.2).
* Version bump
* Add set priority flag so CUDNN_ATTENTION is selected on cuda devices first.
---------
Co-authored-by: Jaret Burkett <jaretburkett@gmail.com >
2026-07-15 12:44:34 -06:00
Jaret Burkett
b8f8a08ba4
Fix sampling bar with anima
2026-07-15 12:17:06 -06:00
rmatif
3e6bd874c4
feat: Add Anima support ( #860 )
...
* Add Anima training support
* Update Anima modular training
* Use sample guidance for Anima
* Fix Anima sampling
* Limit Anima LoRA targets
* Convert Anima LoRA exports
* Fix Anima local loading
* Update Anima default model
* Pin upstream Anima diffusers
* Adjust template defaults to be consistent with other models. Update README
---------
Co-authored-by: Jaret Burkett (Ostris) <jaretburkett@gmail.com >
2026-07-15 11:59:01 -06:00
Jaret Burkett
28f2c0acbe
Move z_image over to the new modeling class
2026-07-13 17:12:30 -06:00
Jaret Burkett
7602e476eb
Exclude sensative layers from quantization in krea
2026-07-13 10:14:53 -06:00
Jaret Burkett
ad07b06de5
Use cached first frame for wan22_5 model
2026-07-10 10:21:34 -06:00
Jaret Burkett
886c2aec57
Allow for vae tiling onle without low vram on wan models with a model kwarg
2026-07-10 09:42:52 -06:00
Jaret Burkett
fe82487187
Add tiling on vae decode for qwen image models when low vram flag is on
2026-07-10 09:07:29 -06:00
Jaret Burkett
e7951ad29e
Exclude sensative layers from quantization on wan models
2026-07-10 08:56:05 -06:00
Jaret Burkett
883d60eb71
Do vae tiling when decoding wan models with low vram active.
2026-07-10 08:02:32 -06:00
Jaret Burkett
a4bbe167ce
Added reference token attention isolation (kv_cache) for Krea2 edit training. Same training cost with significant inference speed up. 2x inference speedup.
2026-07-09 12:00:55 -06:00
Jaret Burkett
6233efe1bb
Fix sampling issue on zimage turbo and krea2 turbo models
2026-07-08 12:34:39 -06:00
Jaret Burkett
7bceec3b07
Apply a loss mask for first frame conditioning for wan and ltx
2026-07-07 09:05:53 -06:00
Jaret Burkett
bd93a312bc
Load more types of comfy ui style quants for ltx2
2026-07-07 09:04:51 -06:00
Jaret Burkett
b1e1a834d4
Added support to train Krea2 as an edit model
2026-07-04 09:12:39 -06:00
Jaret Burkett
7ee1f98f6d
Allow loading and saving z_image in comfy safetensors format.
2026-06-28 07:15:49 -06:00
Jaret Burkett
f8667f0334
Save Krea2 even if quantized
2026-06-28 06:03:37 -06:00
Jaret Burkett
4a99ddabad
Fix breaking change with diffusers qwen image
2026-06-26 10:04:35 -06:00
Jaret Burkett
5f04ae7ad5
Rework merge_network_on_save to handle dequantization on merging and saving for much more efficient full finetuning.
2026-06-25 13:19:38 -06:00
Jaret Burkett
7a089fd0d7
Add support for training directly on Krea2 Turbo with a training adapter
2026-06-23 20:32:02 -06:00
Jaret Burkett
a803611ec1
Enable tiling on vae when decoding with low_vram flag on krea2
2026-06-23 18:38:52 -06:00
Jaret Burkett
724e67d634
Set krea 2 to use new lokr format
2026-06-23 13:31:50 -06:00
