Description

4-step MOSS-Image in 3.4 MB. Two few-step LoRAs for MOSS-Image SFT, trained from zero: 4 steps at CFG 1 instead of 25 steps at CFG 2.5.

- Turbo: the most robust on text (layout JSON 6 / 6, a 16:9 three-line poster 3 / 3 in our tests).

- Turbo Diverse: keeps more of SFT's own variety on open prompts (faces, poses, props).

Settings: 4 steps · CFG 1 · euler / simple · strength 1.0. These are the only settings we measured.

Easiest in ComfyUI: install HaDeS MOSS Suite 0.2.1 from the ComfyUI Manager, plug a MOSS Turbo LoRA node into the turbo input of MOSS KSampler, and it applies the LoRA with those settings for you. The suite ships a ready Turbo example workflow. Stock nodes work too: LoraLoaderModelOnly + KSampler with the settings above.

Measured, not guessed. Every number (text accuracy, layout, 2048 x 2048, aspect ratios, seed variety, ethnic spread on open prompts) is in the model card and in the suite's RESULTS.md, including where the LoRAs fall short (small text at 2048 x 2048).

- Hugging Face (files, model card, full measurements): https://huggingface.co/DeVilDoNia/HaDeS-MOSS-Turbo

- HaDeS MOSS Suite (nodes, training script, every test): https://github.com/Devildonia/HaDeS-MOSS-Suite

Not tested: other base models (Base, finetunes, merges), other strengths or step counts, img2img quality.

Credits: trained from MOSS-Image SFT by the OpenMOSS Team (Apache 2.0); the DMD training recipe follows akhaliq's MOSS-Image-SFT-4-steps (Apache 2.0), none of its weights used.

HaDeS