Open-Sourced MiniMax-H3 Ref-V2V Workflow: Precise Regional Motion Control
Primary_Internal9365 · reddit · 2026-08-08
Searching for a local, controllable alternative to closed APIs like Kling or Gemini, the author built a Reference Video-to-Video (Ref-V2V) ComfyUI workflow based on MiniMax-H3 and open-sourced a custom node.
Core Techniques & Solutions:
- Custom Attention Mask Node: Hooks into DiT self-attention layers, integrating with SAM3 for region-specific control (e.g., locking character motion while replacing backgrounds).
- Smart 32-Multiple Padding: Dynamically applies Reflect Padding to round input resolution to the nearest multiple of 32, preventing dirty resizing and resolution loss.
- 17-Frame VAE Sync: Automatically matches the temporal VAE packing rules to prevent decoding crashes or abrupt truncations.
- Boundary Drift Fix: Implements a Head & Tail Freeze Frame Padding system (15 dummy frames each) to solve motion warping at the video's end.
The author also highlights an unsolved mystery: the frame index offset required for trimming fluctuates between 15 and 30 across different videos. Tested on an RTX 5090, the workflow supports an unofficial 4-step Turbo LoRA for acceleration.
More from coding & agent
- Constraints you force on models today will become tech debt in months, warns dev — zack_overflow · 2026-10-16
- Liquid AI tutorial: fine-tune your d1-3B decision model in 30 minutes on a free GPU — helloiamleonie · 2026-10-16
- Alibaba open-sources OpenGCUA, a full-stack generalist computer-use agent; +19.6 points on OSWorld-v2 across 8 models — Tongyi-MAI · 2026-10-16
- BotBeat: a social network where all users are AI agents, exposed via 32 MCP tools — Dismal_Piccolo4973 · 2026-10-16
- DKGI paper: agents write their own tools to search knowledge graphs, beating baselines by 6.7 Hit@1 — yuz9yuz · 2026-10-16
- Dev open-sources lgtmaybe: running PR code reviews on a local 27B model via hunk splitting — mattjcoles · 2026-10-16