Moss Image Beauty LoRA

Moss Image Beauty LoRA

Portrait and beauty-photography LoRA for the MOSS-Image models, giving tidy close portraits with soft light and fine skin and hair detail, triggered by moss_beauty_style. For image makers using Diffusers or ComfyUI.

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: This is a LoRA adapter that gives MOSS-Image models a portrait and beauty-photography look. It leans toward tidy close-up faces, gently shaped light, fine skin and hair texture, plain backgrounds and magazine-style beauty shots. It works with MOSS-Image-SFT, which the author prefers, or with MOSS-Image-Base. A trigger word at the start of your prompt switches the style on.

Who it suits: It suits image makers who already run MOSS-Image models and want a consistent beauty-portrait style without writing long style prompts.

Industry | Who uses it | How they use it | Example
Photography | Retouchers and photographers | Draft editorial-style portraits and test lighting ideas | Trying a soft side-light look
Beauty and fashion marketing | Campaign teams | Make concept images for a beauty mood board | Building a mood board of close portraits
Digital art | Illustrators | Use the polished portrait style as a base to paint over | Starting a character portrait
Image tooling | Developers | Add one more style adapter to a MOSS-Image workflow | Switching the adapter on in a ComfyUI graph

If you want other styles, or you do not use MOSS-Image models, you can skip it.

Where to use it: You load it on top of a MOSS-Image model.

• Diffusers, using the main weights file
• Native ComfyUI, using the separate ComfyUI weights file
• stable-diffusion.cpp, using the author's separate GGUF adapter
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Beauty portrait from a prompt | Image makers | A 1024 by 1024 portrait in the adapter's style
Compare checkpoints | Researchers | Results from the 2,000-step and 1,750-step adapters
Mix style strength | Artists | Subtler looks with the scale set to 0.7 to 0.9

How to set it up: You need a MOSS-Image-SFT or MOSS-Image-Base model, plus either Diffusers or ComfyUI. For native ComfyUI, use a ComfyUI checkout at commit b1a247c or later, since that includes native MOSS-Image support.

• A MOSS-Image-SFT model, or MOSS-Image-Base as the alternative
• Diffusers with Python, or a recent ComfyUI
• The adapter weights from this repository

1. Download the repository from Hugging Face.
2. For ComfyUI, take the ComfyUI file and place it in ComfyUI/models/loras.
3. Load it with the usual LoRA loader node, setting the model weight to 1.0 and the CLIP weight to zero.
4. Begin each prompt with the trigger word moss_beauty_style.

pip install -U huggingface_hub
hf download axiomofmind/Moss-image-beauty-lora

hf download axiomofmind/Moss-image-beauty-lora \
  MOSS-Image-Beauty-LoRA-ComfyUI.safetensors \
  --local-dir ./ComfyUI/models/loras

How to use it: Put moss_beauty_style at the start of the prompt, keep the LoRA scale at 1.0 and leave the negative prompt empty. The author's tests also found scales of 0.7, 0.8 and 0.9 worked well.

Setting | Recommended value
LoRA scale | 1.0
Inference steps | 30
Guidance scale | 3.0
Resolution | 1024 x 1024
Negative prompt | empty

Tool | What it does
moss_beauty_style | Trigger word that switches on the portrait style
pytorch_lora_weights.safetensors | Main weights, for Diffusers
MOSS-Image-Beauty-LoRA-ComfyUI.safetensors | The same weights, rekeyed for native ComfyUI
alternative-checkpoint-1750 | Optional newer adapter trained on fewer images

To fetch only the alternative checkpoint, use this command.

hf download axiomofmind/Moss-image-beauty-lora \
  --include "alternative-checkpoint-1750/**" \
  --local-dir ./Moss-image-beauty-lora

Example: You write a prompt that starts with moss_beauty_style and asks for a close portrait lit by soft window light, with a plain backdrop. With scale 1.0, 30 steps and a guidance of 3.0, you get a polished 1024 by 1024 portrait in the adapter's look.

For newcomers and experts: A LoRA is a small add-on file that nudges a bigger image model toward one style. Start with the default scale and one short prompt that begins with the trigger word.

For experts, training used MOSS-Image-Base, and adapter_config.json still names it as the parent. Its transformer targets also fit MOSS-Image-SFT, and the SFT pairing gave the author's preferred results. The ComfyUI file carries the same 504 tensors with only the key prefix changed, so values are untouched.

You load a MOSS-Image model
You add the LoRA at scale 1.0
Your prompt starts with the trigger word
The model denoises in 30 steps
A 1024 by 1024 portrait comes out

Good to know: The top-level adapter comes from 2,000 training steps over a set of 176 images. The optional one used 1,750 steps and a narrower set of 149 images. Both use the same trigger word.

Title: Training images per checkpoint
Main checkpoint, 2,000 steps: 176 images
Alternative checkpoint, 1,750 steps: 149 images

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: your prompts and the model files you download
• Not for: loading the Diffusers weights straight into native ComfyUI, which can print key warnings

On Hugging Face, 13 November 2026

• Task: text to image
• Library: diffusers
• Downloads in the last 30 days: 5,546
• Likes: 22
• License: apache-2.0
• Based on: OpenMOSS-Team/MOSS-Image-SFT
• Tags: diffusers, text-to-image, lora, moss-image, portrait, photography

Links

• Model: https://huggingface.co/axiomofmind/Moss-image-beauty-lora
• Files: https://huggingface.co/axiomofmind/Moss-image-beauty-lora/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download axiomofmind/Moss-image-beauty-lora

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