Repository navigation
Releases
1.3
1.2
1.1.1
Fix: stale CUDA graph replay crash in Describe Reference
Fixes a hard crash (Windows access violation / "scatter gather kernel index out of bounds") that could happen in Describe Reference when describing a second image right after a first one.
Cause: ComfyUI's dynamic-VRAM text generation caches a CUDA graph per transformer layer during a generate() call and replays it on the next call as long as the model's weights match — but that check never accounts for the input shape (token count) of the current call. Describing a second image with a different token count than the first could replay a graph captured for the previous call's buffer sizes, corrupting GPU memory and crashing the whole ComfyUI process.
Fix: Describe Reference now drops any cached CUDA graph on the text encoder before every generation call, forcing a fresh capture each time instead of a mismatched replay.
Impact: if you were using --disable-cuda-graphs as a workaround for this crash, you can remove it — the fix is scoped to this node's model only, so the rest of your pipeline keeps the CUDA graph optimization.
1.1
v1.1: multi-image references
New
- Multiple person photos: use several photos of the same person (front, back, close-up…) as identity references. Custom Shots numbers them <Picture 1..N>, and the hair and hairstyle come only from the person references.
- Multiple outfit photos: outfit references can be split across several images (front/back of the top or bottom, shoes, 1–4 accessories). Describe Reference assigns each category to the right image socket and describes only what that image is meant to show.
- Image Aggregator node: merges several images into a single input for the person and outfit sockets.
- Updated example workflow (2-stage sampling + upscale) and new example images in example_results/1.1.
Improved
- Custom Shots prompt: it now says explicitly that hair, hairstyle and haircut come from the person reference and stay identical across all shots. The outfit references are marked as preserving only the outfit, never the hair. This prevents the model from mixing the person's hair with the hair in an outfit photo.
- Describe Reference answer quality: refusals, non-answers ("Nothing else of the outfit is visible…", "None visible."), and overly long or multi-line outputs are detected and retried with a new seed (up to 4 attempts). If every attempt fails, the image is skipped instead of polluting the prompt.
Notes
- The person and outfit photos should be consistent with each other (same hair, same person). Inconsistent inputs produce inconsistent results.
- An outfit image that isn't assigned to any category slot (for example an earring close-up) is described as "nothing else visible". Assign it to an accessory_image_N slot.
1.0
v1.0: reference sheets of a person wearing a specific look, built on MiniMax H3's multi-reference conditioning (ref2va). Two reference photos in — one of the person, one of the outfit — and H3 renders that person, in that look, from every angle the sheet needs.v1.0: reference sheets of a person wearing a specific look, built on MiniMax H3's multi-reference conditioning (ref2va). Two reference photos in — one of the person, one of the outfit — and H3 renders that person, in that look, from every angle the sheet needs.