Blender camera moves + MiniMax H3 ref2vid in ComfyUI: full workflow shared
Time-Ad-7720 · reddit · 2026-09-27
The author shares a workflow for pre-building camera motion in Blender, then feeding it to MiniMax H3's reference-to-video in ComfyUI to control AI video camera paths.
Steps: used Codex CLI with Blender MCP to build a mannequin scene and animate the camera; kept the body planted with head/eyes tracking the lens; rendered a 5-second 1080x1080 30fps reference; loaded the clip with a character sheet and background image into ComfyUI, with prompts assigning each input a role (character, environment, or camera motion).
Settings: RTX 5070 Ti + 32GB RAM, 20 steps (53s/step), 0.6 resolution, 1:1 aspect, no turbo LoRA — about 17m40s sampling time. Blender source resolution and generation resolution are independent.
Caveats: outputs still need checking for identity drift, framing changes, background consistency, and loop seams. Full project files, Blender scene, workflow and prompts are publicly available.
Related event: Workflow Combines Blender Camera Previs with MiniMax H3 Video Generation(2 posts)→
More from coding & agent
- Dev shows off building a retro FPS entirely via Codex Mobile — 503 tests passing — Dimillian · 2026-11-11
- Explosion launches Ellf directory: 52 NLP templates plus multi-agent annotation workflow — _inesmontani · 2026-11-11
- Meme: 'Stochastic AI Agent Architecture' Is Just a Markov Chain in Disguise — PrasoonPratham · 2026-11-11
- OpenCode adds Gemini 4 Carbon and Claude Fable 5.2 support — dean_rie · 2026-11-11
- OpenClaw v2026.11.3 patch fixes cert verification and credential exposure across 56 PRs — dean_rie · 2026-11-11
- Fable 5.2 captured 1 in 4 Claude replies on day one, then growth stalled — dean_rie · 2026-11-11