MOSS-Image-SFT GGUF

MOSS-Image-SFT GGUF

Unofficial GGUF files of the MOSS-Image-SFT diffusion transformer in Q8_0 and Q4_K_S, for running text-to-image locally with stable-diffusion.cpp. Needs a separate Qwen3.5-9B encoder, tokenizer and VAE.

Made by axiomofmind on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: A third party converted MOSS-Image-SFT, a roughly 5B diffusion transformer for text-to-image that was tuned from MOSS-Image-Base, into GGUF files. The files hold only the diffusion transformer. Three other parts come from elsewhere: a Qwen3.5-9B language model as the text encoder, the tokenizer from MOSS, and the shared MOSS and FLUX.2 autoencoder. They are for diffusion runtimes, so Ollama, LM Studio and llama.cpp text chat cannot use them.

Who it suits: It suits people who want to run MOSS-Image locally with smaller files and who are comfortable with a command line tool.

Industry | Who uses it | How they use it | Example
Digital art | Illustrators | Generate concept images on their own machine | Producing a scene from a long text prompt
Product design | Designers | Sketch visual ideas from text without a hosted service | Testing a mood image
Software | Developers | Wire the model into a local image pipeline | Running sd-cli from a script
Research | Hobbyists and researchers | Compare Q8_0 and Q4_K_S output on the same prompt | Checking quality against file size

If you want a text chat model, or a one-click app, you can skip it.

Where to use it: You run it with a diffusion runtime that understands the MOSS-Image format.

• stable-diffusion.cpp, through its sd-cli program
• ComfyUI, only through an experimental community GGUF route
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Local text to image | Creators | A 1024 by 1024 png from a text prompt
Low-memory runs | People with limited VRAM | Output with the CPU offload option on
Quant comparison | Researchers | Q8_0 against the smaller Q4_K_S result

How to set it up: Besides the transformer file, you need three more parts from the MOSS pipeline. The encoder and VAE are deliberately not copied into this repository.

• A Qwen3.5-9B text encoder, for example a Q4_K_M GGUF
• The tokenizer.json file taken from the MOSS release, which you must name in the command, since MOSS has its own vocabulary
• The official MOSS and FLUX.2 VAE file, diffusion_pytorch_model.safetensors

1. Get stable-diffusion.cpp, which includes the sd-cli program.
2. Download the transformer file you want from this repository.
3. Download the encoder, tokenizer and VAE.
4. Run sd-cli with all four files and your prompt.

pip install -U huggingface_hub
hf download axiomofmind/Moss-Image-SFT-GGUF

How to use it: Run sd-cli with the transformer, encoder, VAE and tokenizer files plus your sampling settings. Add the CPU offload flag if video memory is short, and swap in the Q4_K_S file for the smaller size.

Tool | What it does
sd-cli.exe | Runs the image generation from the command line
--diffusion-model | Points to the MOSS transformer GGUF file
--llm | Points to the Qwen3.5-9B text encoder
--offload-to-cpu | Moves parts to the CPU when VRAM is limited

sd-cli.exe `
  --diffusion-model MOSS-Image-SFT-Q8_0.gguf `
  --llm Qwen3.5-9B-Q4_K_M.gguf `
  --vae diffusion_pytorch_model.safetensors `
  --tokenizer tokenizer.json `
  --cfg-scale 3 `
  --steps 50 `
  --sampling-method euler `
  --scheduler discrete `
  --flow-shift 5 `
  --seed 42 `
  --rng cpu `
  --sampler-rng cpu `
  -W 1024 -H 1024 `
  -p "A vast abandoned Art Deco railway concourse transformed into a lush indoor botanical garden, tall arched windows, brass clocks, patterned terrazzo floor, morning sunbeams through mist, tiny visitors for scale, highly detailed architectural photography, natural colors." `
  -o moss-sft-q8.png

Example: You run the command above with the Q8_0 file and a prompt about a ruined railway hall full of plants. It writes moss-sft-q8.png at 1024 by 1024 pixels, using 50 steps and a guidance scale of 3.

For newcomers and experts: GGUF is a compact file format, and a diffusion transformer is the part of an image model that shapes the picture. Start with the Q8_0 file and the sample command, changing only the prompt.

For experts, Q4_K_S is mixed precision. The 42 ffn_down matrices plus four early value projections sit at Q5_K, and embedders, norms and output tensors stay at the precision of the source. The OpenMOSS user guide recommends the flow-matching Euler sampler with a shift of 5.0, and skipping layer 22, counted from zero, in the negative branch.

You pick Q8_0 or Q4_K_S
sd-cli loads it with the encoder and VAE
The tokenizer turns your prompt into tokens
The transformer denoises over 50 steps
A png file is written

Good to know: The official OpenMOSS ComfyUI plugin does not load these GGUF files. An experimental community route exists, but the card says quality parity and every quantization tier are not certified. The official guide suggests about 1M, 1.5M or 2M pixels, with sides divisible by 16.

Title: File size on disk
MOSS-Image-SFT Q8_0: 5.314 GiB
MOSS-Image-SFT Q4_K_S: 3.226 GiB

• License: Apache 2.0, as Hugging Face lists it
• Cost to run: none named
• Data it reaches: your prompts and the model files you download
• Not for: text generation in Ollama, LM Studio or llama.cpp, since these are image model files

On Hugging Face, 13 November 2026

• Task: text to image
• Parameters: 5.1B
• Weights: about 2.5 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 18,762
• Likes: 22
• License: apache-2.0
• Languages: en, zh
• Based on: OpenMOSS-Team/MOSS-Image-SFT
• Tags: gguf, moss-image, text-to-image, image-generation, stable-diffusion.cpp, comfyui

Links

• Model: https://huggingface.co/axiomofmind/Moss-Image-SFT-GGUF
• Files: https://huggingface.co/axiomofmind/Moss-Image-SFT-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download axiomofmind/Moss-Image-SFT-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.