b69b983145
Add an info button on the files to show size, format, and layer keys and shaped for the safetensors files in the ui
Jaret Burkett2026-10-27 10:16:13 -06:00
221cac39a0
Improve unified memory usage on the RTX Spark
Jaret Burkett2026-10-27 08:13:02 -06:00
73ebae13a4
Increase the scale range of DFEs velocity scale correction
Jaret Burkett2026-10-24 10:12:38 -06:00
609f3e5fa5
Fix issue with memory manager on MPS devices
Jaret Burkett2026-10-23 14:37:50 -06:00
4370d60904
Remove the compile for qwen3.8 since the kernels are optimized already
Jaret Burkett2026-10-18 13:30:04 -06:00
85d01c6a04
Prevent possible traversal upstream through caption tags on captioner
Jaret Burkett2026-10-18 13:05:41 -06:00
7c6b168c2c
Add tokens that will pull files with extensions git status
Jaret Burkett2026-10-18 13:04:58 -06:00
e72468b456
Added initial support for fine tuning and captioning with Qwen 3.8 27B
Jaret Burkett2026-10-18 12:25:28 -06:00
8cf43dd2cc
Fix issue that left text encoder on gpu after caching text embeddings.
Jaret Burkett2026-10-17 17:22:42 -06:00
7b94dc5d8d
Add info about the Comfy UI node for the MOSS image edit loras
Jaret Burkett2026-10-17 15:46:40 -06:00
577004a9cd
Publish libdevice as a module global so older triton builds without closure capture can compile the convrot kernels
Jaret Burkett2026-10-17 14:54:32 -06:00
9d408eb64d
Fixed issue with Flux.3 controls where references were not scaled properly and more than one reference could be ignored in certain instances
Jaret Burkett2026-10-17 14:43:02 -06:00
88360c68b2
Speed up MOSS-Image edit kv_cache training: reference isolation as unmasked flash attention instead of a boolean mask
Jaret Burkett2026-10-17 12:10:31 -06:00
9a1828f201
Save qkv lora weights for MOSS Image as unfused LoRA, so they are more compatable with both fused and unfused weights.
Jaret Burkett2026-10-17 09:55:04 -06:00
7045370346
Keep cached pixel tensors as uint8 in the dataloader workers and only materialize to fp16 on the main thread
Jaret Burkett2026-10-14 13:40:49 -06:00
07abdbedc7
Add match_target_res for Qwen Image 2.1 so reference images are scaled to the target's pixel area with their own aspect ratio, consistently across text-embedding caching, training, and sampling, and key the embedding cache on it.
Jaret Burkett2026-09-22 13:57:49 -06:00
f073188300
Fix Qwen Image Edit Plus batch_size > 1 without cached text embeddings by splitting the batch-stacked control tensor into one control image per prompt.
Jaret Burkett2026-09-22 12:14:19 -06:00
a35d833abf
Fix text encoder unload skipping models whose pipeline does not expose the text encoder (Qwen-Image-2.1, Z-Image, LTX2), so cached-embedding runs actually free the TE.
Jaret Burkett2026-09-22 12:08:29 -06:00
543a0d74d0
Fix issue with saving device state when a text encoder is unloaded.
Jaret Burkett2026-09-22 12:03:09 -06:00
8fa15e3569
Move the ui model cards to their own plugin folders so plugins can add custom models in the trainer/generator/ui
Jaret Burkett2026-09-19 12:07:07 -06:00
080873bec9
Harden ace step transcription engine to detect and break out of loops. Add melbanroformer to seperate out the vocals and music for better transcriptions. Much better results.
Jaret Burkett2026-09-17 09:49:19 -06:00
1db6083645
Show more information about LoRAs when loading loras in generate page
Jaret Burkett2026-09-17 09:35:43 -06:00
ef8110cef7
Add LoRA uploader to the lora picker on the generator
Jaret Burkett2026-09-17 05:17:34 -06:00
e65c4d0fb6
Fix issue with Krea2 patch size when training with shift
Jaret Burkett2026-09-16 13:46:39 -06:00
410206e41c
Improve speed of sage sheet decoder with yue2. Recover from tracks that fail
Jaret Burkett2026-09-16 12:24:39 -06:00
ce64e65c39
Add support for Qwen 2.5 Omni and add lora loading for captioning models
Jaret Burkett2026-09-16 10:36:19 -06:00
wenyifan
Jaret Burkett
vdruts