Sign In

[MOSS Image] Hybrid 4 + 8 step workflow with TurboLoRA

Download

1 variant available

Config Other

workflow-ui.json

13.17 KB

Verified:

Type
Workflows
Stats

28

Reviews

No reviews yet

Published

Nov 16, 2026

Base Model
Hash
AutoV299858E00B4
01-flower-shop.png

A two-sampler ComfyUI graph for MOSS Image. The first 4 steps run on the base model with er_sde at CFG 2.5. The last 8 steps run with DiffSynth-Studio's TurboLoRA at 1.0 with euler at CFG 1. Both stages sit on one 12-step grid with the same seed and shift 5.

Measured on an RTX 5080 16 GB with ComfyUI 0.42.1, the int8 DiT and about 1 MP: 4.35 s warm and 5.6 s cold, against 9.3 s for our 25-step default. It is not the fastest option. TurboLoRA alone took 2.20 s at 8 steps, but it fully kept the medium in 0 of 6 painted cells, and in 2 of 6 at 16 steps. The 25-step default kept all 6. The base steps at the start are what bring the look back toward the default.

This is not our default. Side by side, we found the 25-step images more natural. Hands and new small artifacts did not pass our review. We chose this graph on one test set of 19 prompts with 2 seeds each, so other seeds and prompts are untested. It is also untested on a 16 GB RTX 4060 Ti.

Files from Comfy-Org/MOSS-Image on Hugging Face: moss_image_sft_int8_convrot.safetensors, qwen3.5_9b_w4a8_gptq.safetensors and flux2-vae_small.safetensors (Apache-2.0). The LoRA is DiffSynth-Studio/MOSS-Image-SFT-TurboLoRA on ModelScope (model-8steps.safetensors, about 621 MB, Apache License 2.0).

The TurboLoRA needs one fix. As published, ComfyUI 0.42.1 maps 0 of its 524 keys, so it would do nothing. A short script adds the missing prefix and writes moss_turbolora_comfy.safetensors, the name this graph loads. The script and the full test are in our article: https://civitai.com/articles/38059

The watercolor, oil and gouache images were prompted as those media.