Moss Image Turbo LoRas

I'll be uploading different Moss Image Turbo LoRAs I use here and crediting the original creators.

Included LoRAs

• lodestones/MOSS-Image-SFT-8-steps-dmd

• lodestones/MOSS-Image-Base-8-steps-dmd

• lodestones/Kromoss 8-steps-dmd2-lora

• akhaliq/MOSS-Image-SFT-4-steps (v2.0)

Intended Use:

• SFT 8-step: Trained for MOSS-Image-SFT, but can also be used on Base.

• Base 8-step: Trained for MOSS-Image-Base.

• Base 8-step DMD2 (from the Kromoss repo): Distilled from MOSS-Image-Base, then refined with a DMD2 GAN stage against real images. Works on Base, SFT and Kromoss. Gives slightly more contrast and detail than the plain DMD version, and is the creator’s recommended pick.

• SFT 4-step v2: Trained for MOSS-Image-SFT. Euler-native, so it runs on the stock sampler, also can be used on Base.

Recommended Settings

Steps:

• 8-step LoRAs (including DMD2): 8 steps. The Base and DMD2 versions also work at 4.

• 4-step LoRA: 4 steps up to 1024×1024, 6 steps around 1536 wide, 8 steps at 2048×2048. At 4 steps above 1024², images look hazy.

CFG: 1.0 (designed to run without CFG)

Sampler: Euler (flow-matching), shift 5

LoRA Strength:

• moss_image_sft_8_steps_dmd_v0.1: 1.0

• moss_image_base_8_steps_dmd_v0.1: 1.0

• moss_image_base_8_steps_dmd2_v0.1: 1.0

• moss_image_sft_4step_v2.0: 1.0

Use one turbo LoRA at a time.

Disclaimer & Attribution

I did not train, distill, or create the original Turbo/DMD LoRA weights. This model page provides modified compatibility versions / repackaged safetensors intended to let compatible LoRA adapters load seamlessly with ComfyUI’s built-in LoRA loader.

Original Creators:

• Full credit goes to lodestones for the MOSS-Image-SFT-8-steps-dmd and MOSS-Image-Base-8-steps-dmd distillations, and the 8-steps-dmd2 LoRA from Kromoss.

• Full credit to akhaliq for the MOSS-Image-SFT-4-steps distillation (warm-started from lodestones’ 8-step work).

License: Apache-2.0

Original Creator Statement / Notes

SFT: “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. Checkpoint: update 2600 (EMA), float32 weights.”

Base: “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. Checkpoint: update 12000 (EMA), float32 weights.”

Base DMD2 (summary): This LoRA continues the 12,000-update Base data-space DMD run for another 3,000 updates (EMA) using DMD2’s second stage. That stage adds a GAN discriminator trained against real images, at a quarter of the learning rate. It is an fp32 rank-128 LoRA. On the creator’s PartiPrompts CLIP benchmark, it scored highest of the 8-step options on both Base and Kromoss. Keep it unmerged rather than merging it into bf16 weights.

SFT 4-step v2.0 (summary): A 4-step, no-CFG distillation of MOSS-Image-SFT. Version 2.0 (released 2026-10-18) is trained directly on the stock flow-matching Euler schedule. It is a single fp32 rank-128 LoRA. The creator reports about 11 s per 1024² image on an A100, accurate English and Chinese text, and no tiling at 2048². For best quality, keep the LoRA unmerged rather than merging it into bf16 weights.