Semi-transparent Canny refs fix pose bleeding in H3 video generation
Vortexneonlight · reddit · 2026-11-02
A Reddit user shares a practical trick for pose-controlled video generation with H3: raw reference images bleed clothes/environment/colors into the video, while pure Canny edges get ignored or hallucinated on complex poses. Overlaying the original image at low opacity on the Canny map ("semi-transparent Canny") makes the model understand both the pose and the elements.
Benefits:
- Less bleeding into the video
- More output variety, less strict prompting
- Enables complex/new poses the model doesn't know
The post includes a full shot-by-shot prompt template (Visual Style, Global Constraint, Asset Mapping, Chronological Actions) with settings: 6 steps, Dmad-Larry dareties mix LoRA, MP 0.7. Attached images compare Canny vs normal no-bg references on acro-yoga poses.
Related event: Semi-transparent Canny trick tackles reference bleeding in video generation(2 posts)→
More from Multimodal
- Restoring Old Scanned Photos with ComfyUI: Which Image Models Work? — GinJockette · 2026-11-02
- H3 Continuity Studio: Open-Source ComfyUI Pack Adds Scene Memory to MiniMax H3 — rambrogi · 2026-11-02
- User Praises Krea 2 for Finalizing Fantasy World Designs — OfficalRingmaster · 2026-11-02
- ServiceNow releases open Apriel Voice 27B: speech, reasoning and tool use in one model — ServiceNow-AI · 2026-11-02
- Akatz Labs releases open-weight Krea 2 OpenCast LoRA for stylized character creation — PurzBeats · 2026-11-02
- 'One Breath': AI-Generated Short Film Wows Reddit With Near-Photographic Realism — Pnarpok · 2026-11-02