MiniMax H3 ref2video extremely slow: ~150s per step on RTX 5070 Ti vs ~27s for image2video
Most-Trainer-8876 · reddit · 2026-09-13
A Reddit user running MiniMax H3 ref2video (int8, Turbo LoRA enabled) in ComfyUI on an RTX 5070 Ti 16GB + 64GB RAM reports 150 seconds per step for a 10s 9:16, 0.2MP video with single image+video input, versus only 27s per step for image2video at 10s 0.5MP. They ask whether the massive speed gap is expected or a misconfiguration, seeking troubleshooting guidance.
More from Multimodal
- Running MiniMax H3 on AMD GPUs: full ComfyUI + ROCm 7.14 Ubuntu guide — Apprehensive_Sky892 · 2026-09-13
- Spawn's AI Environment Design Praised as 'Next Level' — TAbrodi · 2026-09-13
- Pixal3D adds multiview input, natively integrated into ComfyUI for local 3D generation — petewoodbridge · 2026-09-13
- 8-minute AI film made entirely with Seedance 2.5 shows one person can make a movie — minchoi · 2026-09-13
- AI filmmakers are filmmakers: a 3D filmmaker's ComfyUI-powered rebuttal — Delphoi_Studio · 2026-09-13
- New Midjourney Style Exploration: Fresh Looks from a Creator — Salmaaboukarr · 2026-09-13