2-step LoRA for H3 image model hits 8MP without CFG, near baseline at 4 steps

_akhaliq · x · 2026-11-07

A pDMD 2-NFE LoRA (arXiv 2609.35768, rank 128 across attention+FFN of all 50 blocks) gives P-H3-Image a true 2-step mode without CFG.

At 8MP/shift 12, 2 steps are usable and sharp; 4 steps is nearly indistinguishable from the 8-step baseline — recommended for quoted text and flat graphics
Ships as a single bf16 safetensors file plus a ComfyUI-named copy; 2912×2912 renders take 2.9s per image
Stock P-H3-Image at 2 steps is mid-denoise and washed out; the merged LoRA lands composition and detail in the same two evaluations, with a demo Space for 2/3/4 steps

More from Multimodal