MOSS-Image-heal

MOSS-Image-heal

Makes images from text prompts with a finetuned version of the MOSS-Image-Base model, in a Gradio demo. Its output leans toward an anime style.

Made by hugging-apps on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: MOSS-Image-heal is a text-to-image demo built on KBlueLeaf/MOSS-Image-heal, a full finetune of MOSS-Image-Base with a 5.08B transformer in bf16. The finetune tames the largest values in the model's residual stream, so the checkpoint is safer to run in fp16 and bf16. It runs on GPU hardware at Hugging Face and carries an MCP label.

Who it suits: People who like trying new image models, especially anime-style output, and people who study numerical stability in large models.

Industry | Who uses it | How they use it | Example
Design | Illustrators | Draft anime-style pictures from text | Write a tag-based prompt and look at the result
Research | Model researchers | Look at the effect of the healing finetune | Compare the base and healed numbers
Software | Developers | Study a diffusers-style pipeline | Read pipeline_moss_image.py
Education | Students of generative models | See how a text encoder, transformer and VAE fit together | Follow the listed components

People who want photographic results can skip it, because the training data skews toward anime.

Where to use it:

• A browser, by visiting the Space on Hugging Face
• A computer of your own that runs Windows, macOS or Linux and has Python 3.12
• An MCP client, since the Space carries the MCP label, although the source gives no connection steps

Use case | Who | What they get back
Text to image | Illustrators | A picture made from your prompt
Try prompt styles | Hobbyists | Results from tags, captions or plain language

How to set it up:

• Python 3.12 and pip on Windows, macOS or Linux
• The Python packages in requirements.txt
• Gradio 6.31.0, the version the Space uses
• A GPU, since the Space is set to zero-a10g

1. Install the Hugging Face hub package.
2. Download the Space with the hf command.
3. Move into the new folder.
4. Install gradio at the version the Space uses, then the listed packages.
5. Start the app with Python.

pip install -U huggingface_hub
hf download hugging-apps/moss-image-heal --repo-type space --local-dir moss-image-heal
cd moss-image-heal
pip install gradio==6.31.0
pip install -r requirements.txt
python app.py

How to use it: Online, visit the Space on Hugging Face and write a prompt. Your prompt may be a list of tags, tags with a written caption, a GBC scene-graph block, or everyday sentences. On your own computer, run the start command and open the local page Gradio serves.

python app.py

Tool | What it does
app.py | Starts the Gradio interface
pipeline_moss_image.py | Holds the bundled image pipeline
examples | Holds the example prompts
NOTICE.md | Lists the license notices

Example: You enter a prompt of anime-style tags and generate. With the model card's default settings of 16 steps, guidance 3 and shift 5, the app returns one picture in that style.

For newcomers and experts: If you are new, start with one of the example prompts. They come from the model card and the official demo assets, so they show what the model does best.

If you know the field, the transformer comes from the healed repository, while the Qwen3.5-9B text encoder, tokenizer, FLUX.2 VAE and MOSSImageScheduler come from the official base release. The healing cuts the stream RMS by 7.5 times while outputs stay close to the base, and training used about 1M curated danbooru posts with GBC captions.

You write a prompt
The text encoder reads it
The healed transformer denoises
The VAE decodes the picture

Good to know:

Title: Largest residual stream value
Base model: 58411 peak value
Healed model: 3083 peak value

• Defaults: 16 steps, guidance 3 and shift 5, the model card's own sample settings
• Style: outputs skew toward anime
• Hardware: the Space is set to zero-a10g
• License: apache-2 for the model weights, and BSD 3-Clause for the bundled pipeline code
• Cost to run: GPU hardware, no API key or paid service is named
• Data it reaches: the prompts you write and the models it fetches from Hugging Face

On Hugging Face, 13 November 2026

• Likes: 2
• Built with: gradio
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/hugging-apps/moss-image-heal
• Files: https://huggingface.co/spaces/hugging-apps/moss-image-heal/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.