🎨 Sol-H3: 5-Second MiniMax-H3 Video in 1.653 Seconds
NVIDIA published audited benchmarks for the Sol-H3 engine: a 5-second MiniMax-H3 clip (1344x768, 24 FPS) is generated in a median of 1.653 seconds on an 8-card NVIDIA B300 system — 4 DiT passes instead of 50, up to 15.05x speedup. The community is releasing new ComfyUI tools for H3.
🌍 Sparse attention, kernel fusion, and INT8 QKV / FP8 transport bring inference of a video model with audio close to real-time on a single node — the cost of generation drops. The community is moving to working pipelines: character replacement, scenarios, multi-LoRA.
👤 ComfyUI now features Viggle-Animate nodes (character replacement, int8 version 21 GB, optimal 6 steps), a 6-step turbo merge, and accelerated face-refine: 995.5 s → 202.1 s on RTX 4090. Generating based on others' references without consent is a legal risk.
Source 1: https://nvlabs.github.io/Sana/Sol-Engine/Sol-H3 Source 2: https://github.com/Saganaki22/ComfyUI-Viggle-Animate-H3