FastH3-Live hits 22fps: acceleration node benchmarks and the --vram-headroom trick
spartong945 · reddit · 2026-09-11
FastH3-Live v1.2.0 lifts live video generation from 18 to 22fps (91.6% of 24fps), with full benchmarks.
Acceleration nodes (448x448×362 frames, 3 runs): Sol + Spectrum was fastest (−23.8% sampler time, 20.96fps) but ranked last on picture quality; author chose sage + Spectrum (−18.1%, 20.18fps).
Text encoder: swapping int8convrot (1.67s) for qwen3vl32bminimaxh3nvfp4awq (0.7s) saves 1s per clip.
fused-turbo vs own quantized weights: identical size (21GB) and speed, but fused-turbo merges Mystic v2.0 motion smoothing (better picture) and handles both fl2va and ref2va in one file — a real win for live streaming with reference images.
Key gotcha: --vram-headroom 3 is required to stop ComfyUI thrashing weights — keeping 3GB VRAM free lets weights load in whole batches; without it PCIe reads spike to 9-13GB/s as weights get evicted and reloaded repeatedly. Ships a borderless player and 400 new scenes.
More from Multimodal
- I tested 11 video upscaling methods for AI video: local, Topaz and cloud compared — lopas99 · 2026-10-28
- AI Horde adds Moss-Image for free, no-account image generation on volunteer GPUs — Wanabeelee2 · 2026-10-28
- AI band Techhalla reimagines Slipknot's 'Wait and Bleed' as a 1971 soul performance — Mother-Pear-2392 · 2026-10-28
- Inside the Hailuo H3 Max retro-MV workflow: 15-second chunks, one reference image, multi-shot prompts — techhalla · 2026-10-28
- Open-source pipeline turns Seedance AI video into real PBR textures on your 3D mesh — OdinLovis · 2026-10-28
- AI-generated cyberpunk short 'VHS Sirens' shows off consistent characters and neon city detail — brythehuman · 2026-10-28