8-step Turbo LoRA for Qwen-Image-2.1: 6.8x faster, editing quality improves

metalfans · reddit · 2026-11-22

chatrlabs released an 8-step Turbo LoRA for Qwen-Image-2.1 Uncensored Omni, cutting the standard workflow from 35 steps / CFG 5.0 to 8 steps / CFG 1.0. It supports both text-to-image and instruction-based editing with the same LoRA file.

Measured speedup: 54.9s → 8.1s per image (6.8x) on two RTX A6000 48 GB GPUs.

Quality (8-step vs 35-step): prompt adherence and general SFW quality within noise; anatomy 69.4→72.9, image editing 78.9→81.6, identity preservation in human editing 55.8→72.1 — three metrics improved. Known flaw: occasional over-sharpened skin or speckled artifacts.

ComfyUI recipe: LoRA strength 1.0, 8 steps, CFG 1.0, Euler sampler, 1 megapixel, Diffusers sigma schedule via ManualSigmas. Attaches to all 224 adapted modules on BF16/FP8/INT8/NVFP4 builds. Weights, code, and benchmarks are open-sourced.

Related event: Turbo LoRA speeds up uncensored Qwen-Image 6.8x in 8 steps(2 posts)→

Original post →

More from Multimodal

Multimodal channel →