MOSS-Image-SFT-GGUF

MOSS-Image-SFT-GGUF

Runs the MOSS-Image text to image model from compressed GGUF files on CPU, Mac or GPU, through Unsloth Studio or stable-diffusion.cpp. For people who make images on their own computer.

Made by unsloth 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: This is a quantized GGUF copy of MOSS-Image-SFT from OpenMOSS, a text to image transformer with 5.1 billion parameters. Unsloth made the files so the model fits on ordinary hardware. They hold only the transformer, so the VAE, tokenizer and text encoder come from other repositories. The record lists English and Chinese.

Who it suits: It suits people who want to generate pictures locally from a text prompt, from casual users of Unsloth Studio to developers who script the command line.

Industry | Who uses it | How they use it | Example
Design | Designers and illustrators | Generate pictures from text on their own computer | A daytime storefront sign with bold clean letters
Marketing | Small marketing teams | Draft visuals without sending prompts to a hosted service | A quick mock-up of a shop sign
Software | Developers | Run the model from the stable-diffusion.cpp command line | A scripted render with a fixed prompt and size
Research | Model testers | Compare quant levels against the bf16 version | Check Q8_0 against UD-Q4_K_XL

People who want text chat or code help can skip it, since this model only makes images.

Where to use it:

• Unsloth Studio, in its Images tab
• A stable-diffusion.cpp release at master-975 or later, with MOSS-Image support
• CPU, Mac and GPU
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Images inside an app | Non-technical users | Pictures from a prompt in Unsloth Studio, which fetches the extra parts for you
Command line renders | Developers | A png file per command, with size, steps and prompt set by flags
Quality comparison | Testers | Scores for each quant file against the bf16 version

How to set it up: You need a few things first.

• Python and pip for the Hugging Face hub tool
• Unsloth Studio, or a stable-diffusion.cpp build with MOSS-Image support
• Room on disk for the transformer, a text encoder, a VAE and a tokenizer

1. Install the hub tool and download this repository.
2. In Unsloth Studio, open the Images tab, select the MOSS-Image-SFT model and one of this repository's GGUF files. Studio then fetches the missing VAE, tokenizer and text encoder by itself.
3. With stable-diffusion.cpp, download the recommended transformer file, the Qwen3.5-9B Q8_0 text encoder, and the VAE and tokenizer from the FP8 repository.

pip install -U huggingface_hub
hf download unsloth/MOSS-Image-SFT-GGUF

hf download unsloth/MOSS-Image-SFT-GGUF MOSS-Image-SFT-UD-Q4_K_XL.gguf --local-dir moss
hf download unsloth/Qwen3.5-9B-GGUF Qwen3.5-9B-Q8_0.gguf --local-dir moss
hf download unsloth/MOSS-Image-SFT-FP8 vae/diffusion_pytorch_model.safetensors tokenizer/tokenizer.json --local-dir moss

How to use it: In Unsloth Studio you type a prompt in the Images tab. With stable-diffusion.cpp you run sd-cli once per image, pointing it at the four downloaded files.

./sd-cli --diffusion-model moss/MOSS-Image-SFT-UD-Q4_K_XL.gguf \
  --llm moss/Qwen3.5-9B-Q8_0.gguf \
  --vae moss/vae/diffusion_pytorch_model.safetensors \
  --tokenizer moss/tokenizer/tokenizer.json \
  --cfg-scale 3.0 --steps 50 -H 1024 -W 1024 -s 42 \
  -p "a storefront sign that says \"UNSLOTH\" in bold clean letters, daytime, sharp focus" -o moss.png

Tool | What it does
Unsloth Studio | Runs MOSS-Image GGUFs with a bundled engine and starts at 1024x1024, 50 steps, CFG 3, shift 5
sd-cli | The stable-diffusion.cpp program that renders one image per run
--diffusion-model | Points to the transformer GGUF file
--llm | Points to the text encoder file
--vae and --tokenizer | Point to the VAE and tokenizer files
--cfg-scale, --steps, -H, -W | Set guidance strength, step count, height and width
-p and -o | Give the prompt text and the output file name

Example: You ask for a daytime storefront sign that reads UNSLOTH in bold clean letters, with sharp focus, and run the command above. It writes a 1024x1024 image to a file named moss.png in your current folder.

For newcomers and experts: GGUF is a compact packaging of a model so it runs on smaller machines. If you are new, start with Unsloth Studio, since it fetches the extra pieces itself. The card names UD-Q4_K_XL, at 3.42 GB, as the pick for most machines.

The UD files mix precisions across layers. Unsloth picked each layer's type by measuring how sensitive it was, using image generation runs to gather an importance matrix. Quality is scored with LPIPS against the bf16 transformer, where lower is closer. UD-Q4_K_XL scores 0.213 and Q4_K_M 0.294. The card warns that one bf16 rounding change in one layer already moves renders by about 0.041, so small gaps are composition shifts.

Prompt goes to the Qwen3.5-9B text encoder
Transformer GGUF runs the 50 steps
VAE turns the result into pixels
Image is saved as a png file

Good to know:

• License: Apache 2.0.
• Cost to run: no API key or paid service named. You need your own hardware and enough disk space for the extra files.
• Data it reaches: the extra model files it downloads from Hugging Face, and the image files you write to disk.
• Not for: use on its own. These files hold only the transformer, so you also need the text encoder, VAE and tokenizer.

Title: Quant files and how close they are to bf16
File | Size | Note
UD-Q3_K_XL | 2.84 GB | The smallest in the list
Q4_K_M | 3.20 GB | Scores 0.294 LPIPS against bf16
UD-Q4_K_XL | 3.42 GB | The card's pick for most machines
UD-Q5_K_XL | 3.95 GB | Closer to bf16, 0.121 LPIPS
UD-Q6_K_XL | 4.53 GB | Closer again, 0.106 LPIPS
Q8_0 | 5.48 GB | The closest quantized file, 0.054 LPIPS
BF16 | 10.17 GB | The full precision reference

On Hugging Face, 13 November 2026

• Task: text to image
• Library: gguf
• Parameters: 5.1B
• Weights: about 2.5 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 161,886
• Likes: 23
• License: apache-2.0
• Languages: en, zh
• Based on: OpenMOSS-Team/MOSS-Image-SFT
• Tags: gguf, unsloth, quantized, moss-image, text-to-image

Links

• Model: https://huggingface.co/unsloth/MOSS-Image-SFT-GGUF
• Files: https://huggingface.co/unsloth/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 unsloth/MOSS-Image-SFT-GGUF

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