FastH3 loses its speed in ComfyUI when a LoRA is plugged into the workflow
nekrovski · reddit · 2026-10-18
The author hit an issue running FastH3 in ComfyUI with the default template workflow: after adding a LoRA, the model outputs of the Model Attention Backend and Model Sparse Attention nodes lose their connections to the Basic Guider and Basic Scheduler, stripping FastH3 of its speedup.
- The LoRA's model input connects to Load Diffusion Model fine
- The break happens on the path from model output to Guider/Scheduler
- The author is asking the community for a fix
More from Multimodal
- AI film 'The Gifted' wins $2.5M grand prize at Future Vision XPRIZE in LA — Glittering-Trick5543 · 2026-10-18
- OBVPM Workflow for MiniMax H3 in ComfyUI ships update 007 — fa6637 · 2026-10-18
- Midjourney multi-sref mix recipe produces cinematic light-and-motion portraits — michaelrabone · 2026-10-18
- Reddit user seeks video deepface workflow that fits in 8GB VRAM — what-pos · 2026-10-18
- Marc Andreessen Backs $50K AI Film Contest Sloptoberfest, Entries Due Oct 29 — basedjensen · 2026-10-18
- Real-time 3D capture on a Shanghai cruise: Tripo API meets VITURE AR glasses — Scobleizer · 2026-10-18