MOSS-Image-SFT-4-steps

⚡ Demo: MOSS-Image 4-Step Turbo Space

🆕 v2.0 (2026-10-18) — major update, usage changed. The model is now Euler-native: a single lora.safetensors that runs on the stock flow-matching Euler sampler (any LoRA loader, 4 steps, CFG 1) — the custom data-space sampler and the time_embedding.safetensors / dataspace.safetensors files are no longer needed and were removed from the repo root. If you used v1.x adapters, grab the old data-space variant from checkpoint-update-400/ (it keeps its own sampler), or switch to the new one-line loader below. New users: just follow the Quickstart.

The fastest high-fidelity MOSS-Image-SFT: 4 steps, no CFG, one LoRA file, stock sampler. No custom loop, no auxiliary safetensors, no pipeline patches — it drops into any standard LoRA loader.

Both images above were generated by this model in 4 steps, no CFG — "TODAY'S SPECIAL / Flat White $4.50" and 深夜食堂 render correctly.

License: Apache-2.0 — commercial use OK.

Why this one

Claim	Evidence
Only 4-step MOSS-Image that runs on the stock sampler in any LoRA loader — ComfyUI LoraLoaderModelOnly, or diffusers via the bundled apply_lora.py (the custom pipeline has no load_lora_weights), release flow-Euler scheduler, CFG 1	v3 is trained natively on the release schedule (anchors t = 1000/910/717/24), not adapted to it
Best text/numeric fidelity of any few-step MOSS-Image	4-way head-to-head, same prompts + seeds: this model is the only one that renders "62% / 38%" correctly — darask0's rapid-moss and lodestones' lora_latest both garble secondary numbers (see strips below)
Bilingual text rendering (中文 + English)	Signage-lexicon trained v1.1+: "OPEN 24H", menus, infographics, Chinese storefronts
No tiling at 2048²	Verified 1024² / 1536×1024 / 2048² (tiling_probe/)
Works on MOSS-Image-Base too	Apply the same lora.safetensors to the Base transformer — text fidelity preserved, output slightly softer (Base's own look, per OpenMOSS). Verified: base_compat/ (community find by @lvladikov)
Fully packaged	All-in-one repo (pipeline + text encoder + VAE), inference.py CLI, live ZeroGPU demo, training code included

~11 s per 1024×1024 image on one A100 (4 forwards, bf16 base + fp32 LoRA), no CFG, no negative prompt. Distilled via data-space DMD (arXiv 2610.12102), warm-started from lodestones/MOSS-Image-SFT-8-steps-dmd.

⚡ 3-step draft tier — the fastest MOSS-Image that stays clean

25% faster than the 4-step (~4.2 s vs ~5.6 s per 1024² image on one A100) with draft-tier quality that actually holds: correct text, single sharp subjects, clean humans. The middle anchor gives the model a layout-correction step — this is what 2 steps structurally cannot do (2-step documented on the card as a failed experiment, strips included).

Examples — 3 steps, CFG 1, same seeds as the eval strips:

Menu	深夜食堂 street	Infographic	Lantern	Tea tin	Shiba Inu

ComfyUI: put lora_3step.safetensors (or the frozen moss_image_sft_3step_v1.0.safetensors) in models/loras/, load it with LoraLoaderModelOnly (strength 1.0) on the stock MOSS-Image-SFT diffusion model, KSampler = euler / simple / 3 steps / CFG 1.0. Same base files and sampler as the 4-step — just one fewer step. Do not run it at other step counts (trained for exactly 3).

Speed ladder: 3-step draft (~4.2 s) → 4-step final (~5.6 s) → 6–8 steps above 1024². Human examples + a 30-second video: showcase/.

Quickstart

🟢 Easiest — ComfyUI (recommended)

1. Update ComfyUI (needs the MOSS-Image support from 16 Oct 2026 — the Load CLIP node must offer type moss_image).

2. Download 4 files into your ComfyUI folders:

   File	Put in	Source
   moss_image_sft_bf16.safetensors (10 GB)	models/diffusion_models/	Comfy-Org/MOSS-Image
   qwen3.5_9b_int8_convrot.safetensors (10 GB)	models/text_encoders/	Comfy-Org/MOSS-Image
   flux2-vae.safetensors (0.3 GB)	models/vae/	Comfy-Org/MOSS-Image
   moss_image_sft_4step_v2.0.safetensors (1.2 GB)	models/loras/	this repo

3. Drag comfyui/moss-image-4step-turbo.json onto the canvas and hit Queue. The workflow already has the right settings: LoraLoaderModelOnly at strength 1.0, sampler euler / simple, 4 steps, CFG 1.0.

⚠️ Running the default MOSS-Image template (25–50 steps, CFG 2.5–3) with this LoRA gives smudged, overcooked output — the distilled model wants 4 steps and CFG 1.0 (6–8 steps above 1024²). Full-quality option: load the merged fp32 single file moss_image_sft_4step_v2.0_fp32.safetensors in Load Diffusion Model and skip the LoRA entirely (see ComfyUI below).

⚡ Draft mode: swap the LoRA file for lora_2step.safetensors (or the frozen moss_image_sft_2step_v0.1.safetensors) and set 2 steps — ~2× faster previews, then flip back to the 4-step for finals.

🐍 diffusers

Base bf16 transformer + this LoRA, release sampler, 4 steps, CFG 1. Note: the custom pipeline has no load_lora_weights — use the bundled apply_lora.py:

import torch
from diffusers import DiffusionPipeline
from apply_lora import apply_turbo_lora   # ships in this repo

pipe = DiffusionPipeline.from_pretrained(
    "akhaliq/MOSS-Image-SFT-4-steps",          # all-in-one: also loads from base repo
    custom_pipeline="pipeline_moss_image", trust_remote_code=True,
    torch_dtype={"default": torch.bfloat16, "vae": torch.float32},
).to("cuda")
apply_turbo_lora(pipe.transformer)             # fp32 LoRA, unmerged
image = pipe("A neon sign reading '深夜食堂' over a rainy door",
             num_inference_steps=4, guidance_scale=1.0,
             generator=torch.Generator("cpu").manual_seed(1234)).images[0]

Precision: keep the LoRA unmerged in fp32 — the distillation delta is smaller than bf16's rounding step. The fp32 LoRA applied unmerged on the bf16 base (via apply_lora.py) is the evaluated configuration; merging into bf16 degrades fine text first. For maximum quality use the merged fp32 transformer/ instead.

💻 Single-file CLI

python inference.py --prompt "..." --steps 4 --out out.png draft: python inference.py --prompt "..." --steps 2 --lora-file lora_2step.safetensors --out out.png

⚡ Draft mode (3 steps) — full instructions

The lora_3step.safetensors draft LoRA uses the same base model, loader, and scheduler settings as the 4-step — only the step count changes:

Path	How
ComfyUI	Put lora_3step.safetensors in models/loras/, load it in the same workflow, set KSampler steps = 3 (CFG stays 1.0)
diffusers	apply_turbo_lora(pipe.transformer, filename="lora_3step.safetensors"), then pipe(..., num_inference_steps=3, guidance_scale=1.0)
Space	Pick Draft (3 steps) in the Mode radio — weights hot-swap
CLI	add --steps 3 --lora-file lora_3step.safetensors

Use it for prompt iteration and composition scouting, then re-render finals at 4 steps. Do not run it at other step counts (trained for exactly 3).

Why not 2 steps? Two documented attempts (data-scaled DMD, ~$8; photo- critic + spectral, ~$7) both failed to beat the plain 2-step baseline, and the 2-step shows mode-averaging at 2 anchors — blurred/duplicated subjects (strips, eval_3step/). The 3-step's middle anchor gives the model a layout- correction step; it's the draft tier that works. The 2-step file stays in the repo for the record.

Resolution ↔ steps (measured, hires_probe/): the 4-step anchors cover ~4k latent tokens. Above that, give it more steps — same LoRA, no retrain:

Output	Steps
≤1024×1024	4
~1536 wide	6
2048×2048	8

At 4 steps above 1024² the output turns hazy/washed-out; at 6–8 steps it is sharp (verified 2048² portraits and street scenes). A multi-resolution distillation for native 4-step-at-2048 is on the roadmap.

Prompt tip for high-res (thanks @lvladikov): above 1024², longer, detailed prompts hold up markedly better than one-liners — short prompts ("Beautiful woman") are where haze and grid-tiling show first; a full scene description with subjects, setting, lighting and text survives much better at 1536–2048. The root cause is single-resolution training (1024² only); a multi-resolution bucket run is planned.

💡 Prompt length matters at high resolution (community finding, @lvladikov): above 1024², short prompts ("Beautiful woman") go hazy even at 6–8 steps, while long descriptive prompts (subject, clothing, light, lens, setting) stay clean at 4. The model was trained on long scene captions — at ≥1536 wide, write a full sentence, not a noun phrase. Short prompts at high res also increase ghosting/subject cloning; a multi-resolution training run to fix this at the weights level is next.

Single-file CLI: python inference.py --prompt "..." --steps 4 --out out.png

Files

• lora.safetensors — the model. fp32 LoRA, rank 128 / alpha 128 on every block's proj.{q,k,v,o,ffn_up_gate,ffn_down} (kohya keys). Euler-native. Canonical name — always points at the latest release. Frozen versioned copies are also provided so you always know exactly which version you have: moss_image_sft_4step_v2.0.safetensors (identical to today's canonical file). Old releases stay in their checkpoint folders (checkpoint-update-* data-space, euler-update-* Euler).
• lora_3step.safetensors — 3-step draft tier (canonical), with frozen moss_image_sft_3step_v1.0.safetensors. Same sampler settings, just 3 steps, CFG 1 — ~25% faster than the 4-step, and the draft tier that actually works: subjects stay single and sharp (the 2-step's mode-averaging blur/duplication is gone — the middle anchor gives the model a layout- correction step), and primary text renders correctly. Checkpoints: euler3-update-{50..200}/ (u200 = published, trained on 4×A100 DDP in 71 min).
• lora_2step.safetensors — 2-step experiment (kept for the record; see the 2-step ceiling note in Draft mode).
• transformer/ — merged fp32 transformer (LoRA baked in), for max quality with the release pipeline.
• euler-update-{100..600}/ — v3 (Euler) training checkpoints, u600 = published.
• checkpoint-update-{250,500,750,1000}/ — v1.x data-space checkpoints (archived; those need the data-space sampler and their bundled time_embedding/dataspace files — see checkpoint-update-400/).
• inference.py, train_dmd.py, apply_lora.py — inference CLI, training script, and the diffusers LoRA loader helper.
• comfyui/moss-image-4step-turbo.json — ready-to-run ComfyUI workflow.
• moss_image_sft_4step_v2.0_fp32.safetensors — merged fp32 single file (LoRA baked in at full precision) for ComfyUI Load Diffusion Model.
• Text encoder, tokenizer, VAE, scheduler, pipeline_moss_image.py — bundled from the base repo so this repo loads standalone.

ComfyUI

Ready-to-run workflow: comfyui/moss-image-4step-turbo.json — native ComfyUI loaders (no custom nodes): Load Diffusion Model = the Comfy-Org moss_image_sft_bf16.safetensors, LoraLoaderModelOnly = moss_image_sft_4step_v2.0.safetensors (strength 1.0), KSampler = euler / simple / 4 steps / CFG 1.0, CLIP type moss_image.

If output looks smudged/overcooked: you are almost certainly running the default MOSS-Image template settings (25–50 steps, CFG 2.5–3) with the distilled LoRA — set 4 steps and CFG 1.0 (6–8 steps above 1024², see the resolution table). A default template ignoring the LoRA (undistilled model at 4 steps) also produces mush; check the console for lora key not loaded.

For maximum ComfyUI quality use the merged fp32 single file moss_image_sft_4step_v2.0_fp32.safetensors in Load Diffusion Model and skip the LoRA entirely — standard LoRA patching casts the fp32 delta into bf16 weights, which is where fine text is lost (see Precision).

Head-to-head

The 4-step race — animated, same prompt + seed, live timers (lodestones lora_latest 8-step | this model 4-step | rapid-moss 4-step):

This model finishes first (5.6 s vs 11.2 s) and is the only panel with all numbers correct — rapid-moss slips to "48%", lora_latest invents "AUUTERT". Full frames and the Shanghai-street race: launch/.

🎬 Launch video with voiceover (25 s): showcase/launch_video.mp4

Static strips, 4-way, same prompts + seeds (left → right: this model (Euler, 4 steps) | v1.1 data-space 4-step | darask0 rapid-moss 4-step | lodestones lora_latest Euler 8-step):

Scorecard: this model ties the best on headline text and menus, wins numeric fidelity outright (the infographic — both competitors garble secondary numbers), and is the only one that needs no custom sampler. darask0's rapid-moss still produces the liveliest street-scene compositions. All 8 strips: v3_final/.

Training

Method	data-space DMD (arXiv 2610.12102) v1.x; Euler-native DMD v3
Teacher	frozen MOSS-Image-SFT (bf16), CFG 3.0, negative branch skip-layer 22
Fake score	base transformer (bf16) + fresh rank-128 LoRA, 1:1 update ratio
v1.0	1,000 updates from lodestones update-2600 warm start (data-space)
v1.1	+400 updates, signage-lexicon + 3,000 real user prompts (fixed small-English-text class)
v3	600 updates, Euler-native, warm-started from lora_latest; anchors t = 1000/910/717/24
Optimizer	Adam β₁=0, grad clip 1.0, EMA 0.999
Cost	~$14 + ~$6 + ~$9 on 1× A100-80GB

Citation

@misc{mossimagesft4steps2026,
  title = {MOSS-Image-SFT-4-steps: 4-step DMD distillation of MOSS-Image-SFT},
  author = {akhaliq},
  year = {2026},
  howpublished = {\url{https://huggingface.co/akhaliq/MOSS-Image-SFT-4-steps}}
}

Builds on MOSS-Image (Apache 2.0), lodestones' 8-step DMD, and Few-Step Generation via Data-Space Iteration.