Getting Garments Right in MiniMax H3 ref2va Workflows Is Harder Than It Looks
__MichaelBluth__ · reddit · 2026-10-31
A Reddit user struggles with a ref2va (reference-to-video) workflow from the comfyui-obvpm-timeline project (247 GitHub stars; ComfyUI timeline/clip extension nodes for MiniMax H3):
- Using 3 images for character likeness and 1 for clothing, no matter how the garment is described, the output never matches the garment — or the scene;
- Running an LLM in LM Studio to help with descriptions is painful due to constant loading/unloading between models and H3;
- The post asks how to do ref2va properly, and highlights VRAM juggling between local LLMs and video models as a common ComfyUI pain point.
More from Multimodal
- Custom Claude pipeline turns one prompt into auto-produced explainer videos in 9:16 and 16:9 — AIandDesign · 2026-10-31
- Zeal drops 'Watch The People Glow [Slop Cannon Edition]' AI video — zealcaiden · 2026-10-31
- Custom Claude pipeline auto-grows a kids' explainer YouTube channel from one-line prompts — AIandDesign · 2026-10-31
- Ghibli-style fantasy anime short 'The Train to Tomorrow' made with Seedance — Hot-Decision9790 · 2026-10-31
- Unlocking any-order generation in pretrained autoregressive image models with future-aware embeddings — Rishav Pramanik · 2026-10-31
- Compositor Open-Sources MatteDesk: Free ComfyUI-Based Shot-Level Alpha Matting Pipeline — fishfinger44 · 2026-10-31