Commit Graph
324 Commits
Author SHA1 Message Date
Jaret Burkett affa411edc Fixed an issue where Flux.2 model VAE can be left offloaded to CPU when encoding control images while caching latents 2026-03-29 09:49:10 -06:00
Jaret Burkett 6a1fc54779 Add t0 loss target 2026-03-28 13:35:21 -06:00
abionda-sc 4ef5cbe5bc Fixing bug where width and height are inverted for control image resizing (#707) 2026-03-28 13:00:32 -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
Jaret Burkett dfde30f231 Fix issue with ltx2 custom te repo path 2026-03-25 09:50:18 -06:00
Jaret Burkett 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
Rayane 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
Jaret Burkett 295094b4b5 Fixed new breaking change in diffusers with with qwen image 2026-03-23 14:10:55 -06:00
Jaret Burkett 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
Jaret Burkett 561e6f201c Fixed an issue with ltx 2.3 i2v training 2026-03-23 12:41:18 -06:00
Jaret Burkett e91827f9be Change gemma repo to lightricks one that is not gated 2026-03-23 11:00:32 -06:00
Jaret Burkett 253cb31362 Fix issue with video and images with no audio on ltx models 2026-03-22 22:09:23 -06:00
Jaret Burkett 4a3d317e2b Fix issue with using the default text encoder with ltx 2.3 2026-03-22 18:53:59 -06:00
Jaret Burkett 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
Jaret Burkett 7e1fdc3844 Remove the 0.1 floor for amplification 2026-03-22 09:01:58 -06:00
Jaret Burkett 0f075fc45e Adjust signal amplification target. Allow signal amplification strength in config. 2026-03-22 08:30:13 -06:00
Jaret Burkett dcd98dc0d5 Add signal amplification 2026-03-21 07:44:18 -06:00
Jaret Burkett b04c64e0f8 Add a dino version of DFE 2026-03-04 08:20:37 -07:00
Jaret Burkett 57d407cfd4 Add support for training lodestones/Zeta-Chroma 2026-03-01 12:52:29 -07:00
Jaret Burkett 1c74ca5d22 Add audio_loss_multiplier to scale audio loss to larger values if desired. 2026-02-19 11:57:44 -07:00
Jaret Burkett e82cf6eec2 Fixed issue that prevented full fine-tuning of flux2 models when using gradient checkpointing 2026-02-06 16:18:43 -07:00
Jaret Burkett 1ce2428722 Shrink text embeds to max token length for LTX-2. Drastically reduces cached text embedding sizes 2026-01-28 12:54:49 -07:00
Jaret Burkett a6da9e37ac Add support for FLUX.2 klein base models 2026-01-17 17:46:25 -07:00
Jaret Burkett 0efed794b4 Fix issue where flux2 would ignore single control image on training 2026-01-17 20:26:35 +00:00
Jaret Burkett e40d7ac605 Ignore i2v on ltx is training on images 2026-01-14 18:46:27 -07:00
Jaret Burkett 9848de7946 Fix issue with ltx cached latents if there is no audio. 2026-01-14 17:27:01 -07:00
Jaret Burkett 73dedbf662 Do caching of latents, first frame and audio when caching latents for LTX2 2026-01-14 11:05:23 -07:00
Jaret Burkett 64fe29b182 Support img 2 vid training for ltx-2 2026-01-13 19:04:56 -07:00
Jaret Burkett 5b5aadadb8 Add LTX-2 Support (#644)
* WIP, adding support for LTX2

* Training on images working

* Fix loading comfy models

* Handle converting and deconverting lora so it matches original format

* Reworked ui to habdle ltx and propert dataset default overwriting.

* Update the way lokr saves to it is more compatable with comfy

* Audio loading and synchronization/resampling is working

* Add audio to training. Does it work? Maybe, still testing.

* Fixed fps default issue for sound

* Have ui set fps for accurate audio mapping on ltx

* Added audio procession options to the ui for ltx

* Clean up requirements
2026-01-13 04:55:30 -07:00
Jaret Burkett 0d5c181843 Fixed issue where the control images would sometimes be ignored on qwen_image_edit_2511 2025-12-30 16:40:49 +00:00
Jaret Burkett e6c5aead3b Fix issue that prevented ramtorch layer offloading with z_image 2025-12-02 16:14:34 -07:00
Jaret Burkett 4e62c38df5 Add support for training Z-Image Turbo with a de-distill training adapter 2025-11-28 08:08:53 -07:00
Jaret Burkett 01cf480233 Add FLUX.2 official weights 2025-11-25 08:52:19 -07:00
Jaret Burkett dadbeda197 Update test weights 2025-11-23 10:51:50 -07:00
Jaret Burkett af8e9ea149 Add initial support for FLUX.2 2025-11-18 11:17:38 -07:00
Jaret Burkett 2e7b2d9926 Added Differential Guidance training target 2025-11-10 09:38:25 -07:00
Jaret Burkett 6f308fc46e When soing guidance loss, make CFG zero an optional target instead of a forced one. 2025-11-04 09:16:15 -07:00
Jaret Burkett c984369294 Fixed resizing of control image resolution for Qwen Image Edit 2509 when using match_target_res 2025-10-30 06:30:01 -06:00
Jaret Burkett 42e5e3cd1c Adjust DFE to handle 5 dimension latent spaces 2025-10-27 07:48:44 -06:00
Jaret Burkett 80418209b8 Fixe a variable that could nt be declared when doing blank prompt preservation 2025-10-23 16:19:56 -06:00
Jaret Burkett ee206cfa18 Added blank prompt preservation 2025-10-22 14:55:13 -06:00
Jaret Burkett ff14cd6343 Fix check for making sure vae is on the right device. 2025-10-21 14:49:20 -06:00
Jaret Burkett 76ce757e0c Added initial support for layer offloading wit Wan 2.2 14B models. 2025-10-20 14:54:30 -06:00
Jaret Burkett b7f85928f3 Fix issue with chroma when not quantizing 2025-10-19 12:13:05 -06:00
Jaret Burkett 91b87e06a1 Reordered logs 2025-10-15 09:15:06 -06:00
Jaret Burkett 7e34a03113 Added queing system to the UI 2025-10-14 12:00:42 -06:00
Jaret Burkett 0c9e1c3deb Fixed some fringe cases for qwen image edit. 2025-10-13 17:10:46 +00:00
Jaret Burkett e9c4d94256 Allow for matching target resolution with control images for Qwen Image Edit 2509 2025-10-10 14:24:27 -06:00
Jaret Burkett 1bc6dee127 Change auto_memory to be layer_offloading and allow you to set the amount to unload 2025-10-10 13:12:32 -06:00
Jaret Burkett 8068755b0a Fixed issue with wan 2.2 getting stuck on CPU 2025-10-09 17:24:25 +00:00