MOSS-Image FP16-fixed

MOSS-Image FP16-fixed

The MOSS-Image Base and SFT image transformers with their weights rescaled so they run in fp16 without overflow, giving the same outputs. For diffusers, ComfyUI and KohakuMoss.

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: This is MOSS-Image, in Base and SFT versions, with one edit inside the weights: the running sum that carries values from block to block is divided by 16. fp16 then has plenty of room to spare, where before it had only about 5 percent.

Per the card, results match the official model apart from rounding, and track the fp32 reference more closely than the stock bf16 files. Only the transformer changes. Everything else is unchanged and official. That covers the text encoder and tokenizer, plus the VAE, the scheduler and the pipeline code.

Who it suits: People who generate images with MOSS-Image in diffusers or ComfyUI and want fp16 to work safely, especially at large sizes or with LoRAs and fine-tunes.

Industry | Who uses it | How they use it | Example
Design | Image generators in ComfyUI | Run fp16 on a larger canvas | A 2560 by 2560 render that no longer turns black
Software | Pipeline developers | Use the diffusers transformer in fp16 | Serving images from a GPU with fp16 support
Research | Model tinkerers | Add LoRAs or fine-tunes on top | Training an adapter without overflow

You can skip it if you already run bf16 or fp32 without trouble, or if you want a different image model.

Where to use it: The files target three places, and the card lists which file goes where.

• diffusers, through the MOSSImageTransformer folders for Base and SFT
• ComfyUI, through single-file fp16 checkpoints
• KohakuMoss, the author's trainer and runner

Use case | Who | What they get back
Safe fp16 generation | ComfyUI users | Images without NaN or black output at large sizes
Diffusers pipelines | Developers | An fp16 transformer matching the official config
Tuning on top | Trainers | More headroom when adding LoRAs or fine-tunes

How to set it up: You need Python and the Hugging Face tools to fetch the files, then one of the three runners. The card says the text encoder and VAE come from the official or Comfy-Org releases.

• Python with pip
• diffusers, ComfyUI or KohakuMoss
• The official text encoder and VAE, which this repository does not change

1. Install the Hugging Face hub tools with pip.
2. Download the repository files with the hf command.
3. Point your runner at the Base or SFT transformer, or the matching ComfyUI file.

pip install -U huggingface_hub
hf download KBlueLeaf/MOSS-Image-FP16-fixed

How to use it: Choose Base or SFT, then load the matching file in your tool. In ComfyUI use the single-file checkpoint, and in diffusers load the transformer folder. The card shows no run command.

File | Used for
base/transformer and sft/transformer | The diffusers MOSSImageTransformer, in fp16
moss_image_base_fp16_fixed.safetensors | A single ComfyUI file for Base
moss_image_sft_fp16_fixed.safetensors | A single ComfyUI file for SFT
reports folder | The numbers behind the page, per image

Example: You load the fixed Base file in ComfyUI with fp16 and render a 2560 by 2560 image. The card says the official file turns black at that size, while the fixed weights stay finite.

For newcomers and experts: If you are new, this is the same image model, adjusted so the numbers inside it stay small enough for fp16. Swap it in where you used the official file and render a normal image first.

If you know the field, the residual stream peaks near 62,000 at block 26 of 42, just under fp16's limit of 65,504. Scaling by 1/16 inside the weights gives 16.9 times headroom there. The card measured 189 forwards and real 50-step sampling to show this.

Official weights peak near the fp16 limit
Scale the residual stream by 1/16
Run in fp16 with wide headroom
Same images up to rounding

Good to know: Plain fp16 on the official file passed every test the author ran up to 2048 by 2048, then failed at 2560 by 2560 on step 29 of 50. The fixed weights keep much more margin.

Title: Headroom to the fp16 limit at block 26, times
Official weights: 1.05 times
Fixed weights: 16.9 times

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the model files you download and the prompts you give your own runner
• Not for: changing the text encoder, VAE or scheduler, since only the transformer is changed

On Hugging Face, 13 November 2026

• Task: text to image
• Library: diffusers
• Formats: Safetensors
• Downloads in the last 30 days: 0
• Likes: 38
• 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-FP16-fixed
• Files: https://huggingface.co/KBlueLeaf/MOSS-Image-FP16-fixed/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download KBlueLeaf/MOSS-Image-FP16-fixed

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