MiniMax H3 Ref-to-Video tested locally: two images work, image+video just clones the input
ButterscotchNo6873 · reddit · 2026-10-12
A Reddit user tested MiniMax H3's "Ref to Video" feature locally in ComfyUI:
- Using two reference images ran successfully, but the output was not particularly convincing.
- Switching to one reference image + one reference video failed entirely: after 45 minutes of local inference, the model output was identical to the input reference video, essentially cloning it.
The author followed the official prompt instructions and suspects a missing component or wiring issue in the workflow (screenshot shows a French-language ComfyUI graph with VAE Decode and MiniMax H3 Ref-to-Video nodes). They are asking the community for help troubleshooting.
More from Multimodal
- AI-generated motion video demo set to music — apostraphi · 2026-10-12
- Redditor Builds AI Workflow to Retexture MakeHuman Characters in Blender, Part by Part — o0ANARKY0o · 2026-10-12
- Another scene drops from AI-animated GTA-style series Sons of Los Santos — therealyungxic · 2026-10-12
- Fixed ComfyUI workflow for Qwen 2.1 using recommended sigmas cuts artifacts — Friendly-Fig-6015 · 2026-10-12
- Have we fully explored SDXL? New experiments suggest untapped potential — PurzBeats · 2026-10-12
- Suno v6 Criticized as Too Formulaic, Lacking the Weirdness of Earlier Versions — Kyrannio · 2026-10-12