ComfyUI-SPEED generalizes into a drop-in KSampler replacement, 1.58x faster video gen without retraining
antipode_insights · reddit · 2026-11-03
- The author generalized SPEED (low-res sampling in the noisy early diffusion stages, stepping up to full resolution as detail emerges) into a drop-in KSampler/CustomAdvanced replacement, tested on MiniMax-H3, MOSS and Krea 2; also works with fibo 2 and flux 3.
- Benchmarks (same seed/prompt): H3 5s 0.5MP clip 288s → 182s (1.58×); MOSS Image 22s → 14–18s; Krea 2 283s → 183–215s.
- Ships example workflows with harvested values, per-input tooltips, CPU-drawn transition noise for cross-GPU seed consistency, deprecated-node compatibility.
Related event: ComfyUI-SPEED: Training-Free Sampling Speedup for Diffusion Pipelines(2 posts)→
More from Multimodal
- One style, endless characters: AI-generated 3D figures impress — OVolosin82152 · 2026-11-03
- Reshoot 4D v0.1: turn any video into a 4D Gaussian Splat and reshoot the camera, open weights on a 3090 — Scobleizer · 2026-11-03
- 7-step MCP for Blender workflow lets Claude build photorealistic scenes end to end — sidahuj · 2026-11-03
- AI filmmaking's spatial continuity problem: LTX 2.5 invents rooms when the camera moves — iJohnnyRobert · 2026-11-03
- After 19 years in game dev, a 5-week Opus 5.5 + Unity 6 experiment made it fun again — Delicious-Shower8401 · 2026-11-03
- Trained an SDXL LoRA for Chinese gongbi portraits, then abandoned it — xiaoxuanchuang · 2026-11-03