MOSS Image with 8-step Turbo LoRA tested: 10x faster with minimal quality loss
Aggressive_Collar135 · reddit · 2026-10-20
A Redditor benchmarked MOSS Image SFT (bf16 + Qwen int8 + Flux2 VAE) with and without the 8-step DMD Turbo LoRA:
- A: 50 steps + euler + sgmuniform; B: 8 steps + turbo LoRA at strength 0.8 + euler + simple
- Result: the Turbo LoRA is roughly 10x faster; both variants still produce occasional body horrors and bad text, though less frequently with the LoRA
- Quirk noted: oil-painting prompts get the general style right, but additional in-image details don't adhere to the styling
- The author ran without the original workflow's flow sigma shift / uncond skip nodes and plans to test JSON layout control next
Both MOSS-Image (Comfy-Org) and the 8-steps-dmd LoRA (lodestones) are available on Hugging Face.
More from Multimodal
- Stanford's Halcyon: strong text-to-image without language pretraining — StanfordUniversity · 2026-10-20
- Asked GPT-6 to show polar bear vs. tiger size — it nailed the proportions — throwaway_anonymous7 · 2026-10-20
- Hyper-real candid morning couple photo: full Midjourney prompt shared — tisch_eins · 2026-10-20
- Helm: training-free activation steering for ComfyUI packs model concepts into ~22KB steerable axes — leolambertini · 2026-10-20
- Z-Image Turbo runs locally on Mac Studio: 8-step CFG-1 recipe yields striking B&W portraits — braddown · 2026-10-20
- Concatenative neural speech synthesis project shared — kastnerkyle · 2026-10-20