MOSS-Image-heal by KBlueLeaf

MOSS-Image-heal by KBlueLeaf

A fine-tune of the MOSS-Image-Base text to image model by KBlueLeaf that shrinks the huge values in its internal stream while keeping outputs close to the base. Suits ComfyUI and diffusers users and numerics researchers.

Made by KBlueLeaf 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: MOSS-Image-heal is a full fine-tune of MOSS-Image-Base, a text to image model, made by KBlueLeaf. The tuning shrinks the extreme values in the transformer's residual stream. The largest one drops from 58,411 to 3,083, well under the fp16 limit of 65,504. Only the transformer was trained. The other pipeline parts, such as the VAE and the text encoder, come unchanged from MOSS-Image-Base.

Who it suits: Image generation developers and ComfyUI users who want a MOSS-Image variant with tamer internal values, and researchers who study large activations.

Industry | Who uses it | How they use it | Example
Creative software | ComfyUI users | Load a single file in bf16, fp16 or int8 | Use the fp16 file in a workflow
Software | Diffusers developers | Load the transformer folder with diffusers | Swap it into the MOSS-Image pipeline
Research | Numerics researchers | Study how the stream values were tamed | Compare the base and heal statistics

Skip it if you want a model that looks clearly different from the base, because the card shows outputs stay close to it.

Where to use it:

• Python with diffusers, using the transformer folder
• ComfyUI, using the single file bf16, fp16 or int8 convrot versions

Use case | Who | What they get back
Text to image with diffusers | Developers | An image made from a prompt
Image generation in ComfyUI | Artists | Images from a workflow that loads the single file
Numerics study | Researchers | Reports holding every number and chart value in the card

How to set it up:

• Python and the Hugging Face tools
• The other pipeline parts from MOSS-Image-Base, such as its VAE and text encoder
• About 10.2 GB of disk for the BF16 weights, as an estimate

1. Install the Hugging Face hub package.
2. Download the repository files.
3. Load the transformer with diffusers, or place a single file in ComfyUI.

pip install -U huggingface_hub
hf download KBlueLeaf/MOSS-Image-heal

How to use it: Use it like the base model, with this transformer swapped in. The card's samples used a caption prompt of tags plus a sentence, and a GBC prompt, with the same negative prompt each time.

Tool | What it does
transformer folder | The diffusers transformer in bf16
ComfyUI bf16 file | A single file with fused qkv
ComfyUI fp16 file | A single file in fp16
ComfyUI int8 file | A single int8 ConvRot file with group size 256
reports folder | Holds every number and chart value in the card

Example: Generate with the card's sample settings: 16 steps, CFG 3, shift 5 and seed 0, using a caption prompt. The picture should look close to what the base model makes with the same inputs.

For newcomers and experts: If you are new to this, try the fp16 single file in ComfyUI first, with the same prompt you use on the base model, and compare the two pictures.

If you know the field, the card reports a stream RMS 7.5 times lower and headroom to the fp16 maximum of 21.2 times, against 1.12 times for the base. No weight now sits above 64 times its tensor's RMS, against 349 before. Flow loss moves from 0.4280 to 0.4289, and the velocity cosine to the base is 0.9983.

The prompt is encoded by the official text encoder
The healed transformer denoises
The official VAE decodes the latent
An image comes out

Title: Largest value in the residual stream
Base: 58411 units
Heal: 3083 units
fp16 maximum: 65504 units

Good to know:

• The sample images in the card put base and heal outputs side by side.
• The card measured the stream statistics on 6 text and image samples at 1024 squared area.
• The base model for this fine-tune is OpenMOSS-Team/MOSS-Image-Base.
• License: apache-2.0.
• Cost to run: none named, only your own hardware.
• Data it reaches: the prompts you give it. The source names no account or outside service.
• Not for: the source names none.

On Hugging Face, 13 November 2026

• Task: text to image
• Library: diffusers
• Parameters: 5.1B
• Weights: about 10.2 GB in BF16; about 2.5 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 0
• Likes: 76
• License: apache-2.0
• Based on: OpenMOSS-Team/MOSS-Image-Base
• Tags: diffusers, safetensors, moss-image, fp16, comfyui, text-to-image

Links

• Model: https://huggingface.co/KBlueLeaf/MOSS-Image-heal
• Files: https://huggingface.co/KBlueLeaf/MOSS-Image-heal/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download KBlueLeaf/MOSS-Image-heal

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