👉 Update (Oct. 28, 2026) - I recommend using these LoRAs with the full official Comfy dot org bf16 versions of SFT and Base, available here:

• MOSS-Image SFT (full bf16) - https://tensor.art/models/1060474182812367660

• MOSS-Image Base (full bf16) - https://tensor.art/models/1060466259671441006

The official TA versions of MOSS-Image models are the official Comfy dot org int8 convrot versions. While fine for most purposes, IMO the bf16 versions offer better prompt adherence and variety.

For Base generations, I also recommend Lodestones' "Kromoss" fine-tune: https://tensor.art/models/1057527871754506597/Kromoss-by-Lodestones-bf16. Lodestones is the legendary creator of Chroma and Kroma, as well as these turbo LoRAs.

**********

From LoRA creator Lodestones (https://huggingface.co/lodestones):

"Usage

load lora.safetensors ["SFT v1" on this TA reprint] (at 1 strength recommended) or lora_latest.safetensors ["SFT v2 ('Latest')" on this TA reprint] (at 0.7 strength recommended) directly to your comfyui lora loader

MOSS-Image-SFT, 8-step data-space DMD

Work-in-progress distillation of: OpenMOSS-Team/MOSS-Image-SFT (50 steps, CFG 3) to 8 network evaluations, no CFG. Data-space DMD (arXiv 2610.12102): the generator iterates on its own clean estimate from fixed noise. Trained on compact-JSON scene_graph captions.

MOSS-Image-Base, 8-step data-space DMD ["Base v1" on this TA reprint]

Work-in-progress distillation of OpenMOSS-Team/MOSS-Image-Base (50 steps, CFG 3) to 8 network evaluations, no CFG. Data-space DMD (arXiv 2610.12102): the generator iterates on its own clean estimate from fixed noise. Trained on compact-JSON scene_graph captions."

NOTE: for the DMD2 version of this turbo LoRA for Base, go here: https://tensor.art/models/1057582086052906158/MOSS-BASE-DMD-8Step