moonbox-ai/MOSS-Image-SFT-MLX

MOSS-Image (SFT) MLX

MLX builds of MOSS-Image (SFT), made by Moonbox for Apple Silicon Macs. Each precision is a folder; keep a folder's files together. Only the parts listed below are converted: the rest of the pipeline comes from the original repository unless the section below says otherwise.

Files

The last column is how far each build is from the original PyTorch model on the same inputs (see the section below for what was measured).

Precision	File	Size	sha256	Checked against the original
bf16	bf16/mflux_model.json	0.4 kB	afa4f3c20233	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/text_encoder/0.safetensors	2.13 GB	9be8575f9978	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/text_encoder/1.safetensors	2.10 GB	76826c2512c5	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/text_encoder/2.safetensors	2.13 GB	d143078bfed0	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/text_encoder/3.safetensors	2.07 GB	86de523b1bb7	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/text_encoder/4.safetensors	2.05 GB	3cf868fb3c34	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/text_encoder/5.safetensors	2.07 GB	90b728c34f6c	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/text_encoder/6.safetensors	2.13 GB	7ab6019f987b	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/text_encoder/7.safetensors	1.19 GB	3e67006774cf	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/text_encoder/model.safetensors.index.json	25.0 kB	d7e73ca28859	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/tokenizer/chat_template.jinja	7.8 kB	a4aee8afcf2e	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/tokenizer/tokenizer.json	20.0 MB	06b9509352d2	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/tokenizer/tokenizer_config.json	1.1 kB	bee8eba30f0e	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/transformer/0.safetensors	2.05 GB	ff269e76dc5d	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/transformer/1.safetensors	2.14 GB	ad1e65cda935	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/transformer/2.safetensors	2.15 GB	961108991d7c	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/transformer/3.safetensors	2.06 GB	9d9d213a41be	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/transformer/4.safetensors	1.78 GB	d255a1280146	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/transformer/model.safetensors.index.json	23.2 kB	d8a42a87e429	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/vae/0.safetensors	124.8 MB	2bc64e7a50c9	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
bf16	bf16/vae/model.safetensors.index.json	16.5 kB	009c1b40da60	the original weights in bf16 (reference for the rows below; not compared with PyTorch)
8bit	8bit/mflux_model.json	0.4 kB	fbcbab761a9e	vs bf16: PSNR 36.0 dB mean, LPIPS 0.011 mean / 0.014 worst
8bit	8bit/text_encoder/0.safetensors	2.12 GB	1d0f2fe9431d	vs bf16: PSNR 36.0 dB mean, LPIPS 0.011 mean / 0.014 worst
8bit	8bit/text_encoder/1.safetensors	2.12 GB	265058f714e6	vs bf16: PSNR 36.0 dB mean, LPIPS 0.011 mean / 0.014 worst
8bit	8bit/text_encoder/2.safetensors	2.12 GB	1c7d9942f7e6	vs bf16: PSNR 36.0 dB mean, LPIPS 0.011 mean / 0.014 worst
8bit	8bit/text_encoder/3.safetensors	2.06 GB	f995a4eba76a	vs bf16: PSNR 36.0 dB mean, LPIPS 0.011 mean / 0.014 worst
8bit	8bit/text_encoder/model.safetensors.index.json	54.2 kB	cfc820f65807	vs bf16: PSNR 36.0 dB mean, LPIPS 0.011 mean / 0.014 worst
8bit	8bit/tokenizer/chat_template.jinja	7.8 kB	a4aee8afcf2e	vs bf16: PSNR 36.0 dB mean, LPIPS 0.011 mean / 0.014 worst
8bit	8bit/tokenizer/tokenizer.json	20.0 MB	06b9509352d2	vs bf16: PSNR 36.0 dB mean, LPIPS 0.011 mean / 0.014 worst
8bit	8bit/tokenizer/tokenizer_config.json	1.1 kB	bee8eba30f0e	vs bf16: PSNR 36.0 dB mean, LPIPS 0.011 mean / 0.014 worst
8bit	8bit/transformer/0.safetensors	2.14 GB	9425d5e4f39f	vs bf16: PSNR 36.0 dB mean, LPIPS 0.011 mean / 0.014 worst
8bit	8bit/transformer/1.safetensors	2.13 GB	9b83123ef826	vs bf16: PSNR 36.0 dB mean, LPIPS 0.011 mean / 0.014 worst
8bit	8bit/transformer/2.safetensors	1.13 GB	c3a91a1c7e35	vs bf16: PSNR 36.0 dB mean, LPIPS 0.011 mean / 0.014 worst
8bit	8bit/transformer/model.safetensors.index.json	50.6 kB	c385aa22daff	vs bf16: PSNR 36.0 dB mean, LPIPS 0.011 mean / 0.014 worst
8bit	8bit/vae/0.safetensors	123.2 MB	2fd063835fd5	vs bf16: PSNR 36.0 dB mean, LPIPS 0.011 mean / 0.014 worst
8bit	8bit/vae/model.safetensors.index.json	17.5 kB	4d629a6165dd	vs bf16: PSNR 36.0 dB mean, LPIPS 0.011 mean / 0.014 worst
6bit	6bit/mflux_model.json	0.4 kB	193edbeac0db	vs bf16: PSNR 28.1 dB mean, LPIPS 0.038 mean / 0.044 worst
6bit	6bit/text_encoder/0.safetensors	2.15 GB	5dc48f61e4ec	vs bf16: PSNR 28.1 dB mean, LPIPS 0.038 mean / 0.044 worst
6bit	6bit/text_encoder/1.safetensors	2.11 GB	a6d44163bb07	vs bf16: PSNR 28.1 dB mean, LPIPS 0.038 mean / 0.044 worst
6bit	6bit/text_encoder/2.safetensors	2.15 GB	fa3e70633240	vs bf16: PSNR 28.1 dB mean, LPIPS 0.038 mean / 0.044 worst
6bit	6bit/text_encoder/3.safetensors	44.1 MB	8d89a3f57f13	vs bf16: PSNR 28.1 dB mean, LPIPS 0.038 mean / 0.044 worst
6bit	6bit/text_encoder/model.safetensors.index.json	54.2 kB	9885f1ccb02a	vs bf16: PSNR 28.1 dB mean, LPIPS 0.038 mean / 0.044 worst
6bit	6bit/tokenizer/chat_template.jinja	7.8 kB	a4aee8afcf2e	vs bf16: PSNR 28.1 dB mean, LPIPS 0.038 mean / 0.044 worst
6bit	6bit/tokenizer/tokenizer.json	20.0 MB	06b9509352d2	vs bf16: PSNR 28.1 dB mean, LPIPS 0.038 mean / 0.044 worst
6bit	6bit/tokenizer/tokenizer_config.json	1.1 kB	bee8eba30f0e	vs bf16: PSNR 28.1 dB mean, LPIPS 0.038 mean / 0.044 worst
6bit	6bit/transformer/0.safetensors	2.12 GB	f6806ca18303	vs bf16: PSNR 28.1 dB mean, LPIPS 0.038 mean / 0.044 worst
6bit	6bit/transformer/1.safetensors	2.01 GB	f758e35a9020	vs bf16: PSNR 28.1 dB mean, LPIPS 0.038 mean / 0.044 worst
6bit	6bit/transformer/model.safetensors.index.json	50.6 kB	2cbc94582e51	vs bf16: PSNR 28.1 dB mean, LPIPS 0.038 mean / 0.044 worst
6bit	6bit/vae/0.safetensors	122.8 MB	bb30345b410e	vs bf16: PSNR 28.1 dB mean, LPIPS 0.038 mean / 0.044 worst
6bit	6bit/vae/model.safetensors.index.json	17.5 kB	a5a21d176b96	vs bf16: PSNR 28.1 dB mean, LPIPS 0.038 mean / 0.044 worst
4bit	4bit/mflux_model.json	0.4 kB	ec9428aa8f27	below threshold (experimental), vs bf16: PSNR 17.0 dB, LPIPS 0.273 mean / 0.292 worst
4bit	4bit/text_encoder/0.safetensors	2.13 GB	3d461b13a81f	below threshold (experimental), vs bf16: PSNR 17.0 dB, LPIPS 0.273 mean / 0.292 worst
4bit	4bit/text_encoder/1.safetensors	2.14 GB	6f8e09165bca	below threshold (experimental), vs bf16: PSNR 17.0 dB, LPIPS 0.273 mean / 0.292 worst
4bit	4bit/text_encoder/2.safetensors	202.9 MB	56de27e47826	below threshold (experimental), vs bf16: PSNR 17.0 dB, LPIPS 0.273 mean / 0.292 worst
4bit	4bit/text_encoder/model.safetensors.index.json	54.2 kB	bd366e569b0b	below threshold (experimental), vs bf16: PSNR 17.0 dB, LPIPS 0.273 mean / 0.292 worst
4bit	4bit/tokenizer/chat_template.jinja	7.8 kB	a4aee8afcf2e	below threshold (experimental), vs bf16: PSNR 17.0 dB, LPIPS 0.273 mean / 0.292 worst
4bit	4bit/tokenizer/tokenizer.json	20.0 MB	06b9509352d2	below threshold (experimental), vs bf16: PSNR 17.0 dB, LPIPS 0.273 mean / 0.292 worst
4bit	4bit/tokenizer/tokenizer_config.json	1.1 kB	bee8eba30f0e	below threshold (experimental), vs bf16: PSNR 17.0 dB, LPIPS 0.273 mean / 0.292 worst
4bit	4bit/transformer/0.safetensors	2.15 GB	ce2b196b2dba	below threshold (experimental), vs bf16: PSNR 17.0 dB, LPIPS 0.273 mean / 0.292 worst
4bit	4bit/transformer/1.safetensors	712.5 MB	147842364c26	below threshold (experimental), vs bf16: PSNR 17.0 dB, LPIPS 0.273 mean / 0.292 worst
4bit	4bit/transformer/model.safetensors.index.json	50.6 kB	1a400badd1dd	below threshold (experimental), vs bf16: PSNR 17.0 dB, LPIPS 0.273 mean / 0.292 worst
4bit	4bit/vae/0.safetensors	122.4 MB	3737818da953	below threshold (experimental), vs bf16: PSNR 17.0 dB, LPIPS 0.273 mean / 0.292 worst
4bit	4bit/vae/model.safetensors.index.json	17.5 kB	e396f8330c32	below threshold (experimental), vs bf16: PSNR 17.0 dB, LPIPS 0.273 mean / 0.292 worst

How to run

Each folder is the whole pipeline (Qwen3.5 text encoder, transformer, FLUX.2 VAE, tokenizer) saved by mflux 0.24.0 from the original weights at the revision below. Run it with mflux:

pip install mflux==0.24.0
hf download moonbox-ai/MOSS-Image-SFT-MLX --include "8bit/*" --local-dir moss
mflux-generate-moss-image --model moss/8bit --base-model moss-image-sft \
  --prompt "A red fox in fresh snow at sunrise, photo" --width 1024 --height 1024 --seed 7 --output fox.png

8bit/, 6bit/ and 4bit/ quantize the transformer's and the text encoder's linear layers (MLX affine, group size 64); the VAE stays bf16. No fp16 folder: mflux computes in bf16 whatever the weights are.

How they were checked. Each quantized pipeline against bf16/ from the same initial noise, 3 prompts at 512x512, mflux's default steps. Pass: LPIPS <= 0.15 on every prompt. bf16/ itself is the original weights in bf16; it wasn't compared with a PyTorch run (diffusers has no MOSS-Image pipeline).

Precision	Size	PSNR vs bf16 (mean)	LPIPS (mean / worst)	Status
bf16/	26.2 GB	reference	-	the original weights
8bit/	14.0 GB	36.0 dB	0.011 / 0.014	passes
6bit/	10.7 GB	28.1 dB	0.038 / 0.044	passes
4bit/	7.5 GB	17.0 dB	0.273 / 0.292	below threshold, experimental

About 2.5 minutes per 512x512 picture on an Apple M5 with 24 GB at mflux's defaults.

How these were made

From MOSS-Image (SFT) at revision f972bd4b4b: the original weights loaded in MLX and stored as bf16 or fp16, or quantized with MLX's affine quantization (8bit, 6bit, 4bit, group size 64; layers MLX can't quantize, such as convolutions and norms, stay bf16). Every build was checked against the original before upload; the sha256 column above is what the files must hash to.

License

Apache-2.0. These files are a derivative of the original model and carry its license: it is copied in this repository, and it governs your use of these files. The original model card follows unchanged.

---

Original model card

MOSS-Image-SFT

                         

User guide   ·   中文用户指南   ·   ComfyUI

MOSS-Image-SFT is a supervised fine-tuned text-to-image model in the MOSS-Image family from OpenMOSS, derived from MOSS-Image-Base. It uses the same architecture and inference pipeline, supporting natural-language prompts, tag sequences, and JSON layouts for controlling text placement and object arrangement.

Base vs. SFT

In this example, SFT produces clearer text grouping and a more organized whiteboard layout.

MOSS-Image-Base	MOSS-Image-SFT

Quick start

Use Python 3.10 or later and an NVIDIA GPU with CUDA support. Install a PyTorch 2.9 or later build compatible with your CUDA environment, then install:

pip install "diffusers>=0.38" "transformers>=5.3" "accelerate>=1.10" "safetensors>=0.6" "Pillow>=10"

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "OpenMOSS-Team/MOSS-Image-SFT",
    custom_pipeline="pipeline_moss_image",
    trust_remote_code=True,
    torch_dtype={"default": torch.bfloat16, "vae": torch.float32},
)
pipe.scheduler = type(pipe.scheduler).from_config(pipe.scheduler.config)
pipe.to("cuda")

prompt = '''A wide view of a refrigerator beside a kitchen counter with a fruit bowl. A small sticky note on the refrigerator door is handwritten in Chinese:
"记得买牛奶和鸡蛋！
牛奶两盒，鸡蛋一打。
顺便带一袋面包。"'''

image = pipe(
    prompt,
    height=1024,
    width=1024,
    num_inference_steps=50,
    guidance_scale=3.0,
    uncond_skip_layers=[22],
    generator=torch.Generator("cpu").manual_seed(42),
).images[0]
image.save("refrigerator_note.png")

The model and custom pipeline code are downloaded and cached automatically. You can also pass a complete local model directory to from_pretrained. The transformer and text encoder use bfloat16; the VAE and initial noise use float32. The bundled VAE uses the FLUX.2 small decoder.

For the Gradio interface, CPU offload, and compilation options, follow the project quick start. For ComfyUI, search for MOSS-Image in ComfyUI Manager to install the native plugin. Integration into the official ComfyUI repository is in progress. See the installation and usage guide and example workflow.

Sampling settings

Setting	Default
Sampler	Flow-matching Euler
Steps	50
CFG	3
Shift	5
Layer skipping	Negative branch, layer index 22 (zero-based)
Negative prompt	Empty

The default guidance uses two transformer forwards per step. Pass uncond_skip_layers=[] to disable negative-branch skipping, or set negative_prompt to provide a negative prompt. Python API image dimensions must be positive multiples of 16.

See the user guide for resolution, sampling, prompt, and LoRA training recommendations.

Prompt enhancer skill

The MOSS-Image Prompt Enhancer turns image ideas into detailed captions or spatial Layout prompts. It is designed for GPT-6-family agents with local file and Python tools. The bundled script converts pixel-coordinate drafts into compact JSON with integer centered coordinates; it does not call an image-generation service.

Install the skill folder in your agent's skills directory, then invoke $moss-image-prompt-enhancer with your image idea, desired caption or Layout output, and any canvas dimensions. Paste the compiled result into MOSS-Image's positive prompt; for Layout, use the same generation dimensions.

Layout-to-image generation

Pass a JSON layout as the prompt to describe the scene, rendered text, objects, and their positions. Layouts use summary.caption and items, with bbox_format: "yxyx" and bbox_coord_space: "abs_centralized".

With the Python API, serialize the layout using json.dumps(layout, ensure_ascii=False) and pass it to pipe. In Gradio, paste the JSON into the Prompt field and set the matching image dimensions. See the layout format and example.

Layouts can also support controlled editing and user-friendly layout tools. Edit text, element descriptions, and bounding boxes to adjust image content and composition in a structured way.

Showcase

Citation

@misc{ai2026mossimagetechnicalreport,
      title={MOSS-Image Technical Report},
      author={MOSI. AI},
      year={2026},
      eprint={2610.23374},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.23374},
}

License and acknowledgements

The model weights and bundled inference code are released under the Apache License, Version 2.0.

We thank the Diffusers, Qwen, and FLUX teams.