MOSS-Image-pose-controlnet

MOSS-Image-pose-controlnet

Makes images follow a body pose by taking a DWPose skeleton as a control image, as a LoRA for MOSS-Image-SFT. Suits artists and animators who want to pose characters in ComfyUI.

Made by thedeoxen 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: A LoRA that adds OpenPose-style pose control to the MOSS-Image-SFT model. You give it a black image with a DWPose skeleton drawn on it. The picture copies that body pose, and your prompt decides who is in it, what they wear and where they stand. It is an edit-style LoRA, not the official MOSS color or depth adapter.

Who it suits: Artists, animators and storyboard makers who work in ComfyUI with MOSS-Image-SFT and want control over body pose.

Industry | Who uses it | How they use it | Example
Illustration | Character artists | Pose a character while changing outfit and style | A full-body anime character in a fighting stance
Photography | Photo editors | Turn the pose of a photo into another person | Reuse a pose with a new face and clothing
Animation | Storyboard artists | Keep the body matched across keyframes | A set of frames sharing one pose

If you do not use MOSS-Image-SFT, you can skip it.

Where to use it:

• ComfyUI, using the workflow file in the repo
• Windows, macOS or Linux, wherever ComfyUI runs
• The DWPose extractor from the comfyui_controlnet_aux nodes

Use case | Who | What they get back
Pose a character | Illustrators | A new character drawn in the pose you supply
Copy a photo's pose | Editors | A different person in the same stance
Storyboards | Animators | Keyframes where the body has to match

How to set it up:

• ComfyUI with the MOSS-Image-SFT model files loaded
• The comfyui_controlnet_aux nodes, for DWPose
• The UNet, text encoder and VAE files the card names

1. Download the repo files, which include the LoRA weights and a ComfyUI workflow.
2. Load the UNet, the text encoder with the moss_image type, and the VAE.
3. Attach the LoRA to the UNet with its strength at 1.0.

pip install -U huggingface_hub
hf download thedeoxen/MOSS-Image-pose-controlnet

How to use it: Make a control map from a photo and hand it to the Text Encode MOSS-Image Edit node as the one reference image. Choose ai-toolkit as the edit format for both the positive and the negative encodes. Run 30 sampling steps at CFG 4, with euler and simple. No trigger word is needed.

Full-body anime illustration of a cyber idol with short white hair and a magenta visor. Upright fighting stance, one arm straight above the head, the other arm extended to the side, one knee lifted to the hip, the standing leg straight. Glass coat, rainy megacity, whole body in frame.

Tool | What it does
DWPose | Extracts the skeleton from a photo, with body, hands and face on
Text Encode MOSS-Image Edit | Takes the control image and your prompt
LoRA strength | Set to 1.0, the trained value, or lower toward 0.7 if the pose feels too rigid

Example: You supply a skeleton in a fighting pose and ask for a futuristic anime singer with white hair, in a glass coat on a rainy street. The result is a full-body picture of that character holding the pose from the skeleton.

For newcomers and experts: If you are new, start with the included workflow. It loads a photo, builds the skeleton, encodes it with your prompt and samples. Describe the character and the world, and let the skeleton carry the pose.

Its adapter ranks are 64 for linear layers and 16 for convolutions. Training treated the task as image editing: the skeleton serves as the reference image and the text describes the output. Training also ran with kv cache on, which is why the edit format must be ai-toolkit. The selected checkpoint is training step 4500.

Load a photo
DWPose builds the skeleton on black
Skeleton and prompt go to the edit encoder
The sampler draws the new picture

Good to know:

• Use a black background, colored bones, and include the face and hands
• Draw every limb in the skeleton, since a prompt does not reliably supply a missing joint
• Your prompt does not need to repeat every joint
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the control image, your prompt and the files downloaded from Hugging Face
• Not for: replacing the official MOSS color or depth adapter

On Hugging Face, 13 November 2026

• Task: image to image
• Library: diffusers
• Downloads in the last 30 days: 813
• Likes: 17
• License: apache-2.0
• Languages: en
• Based on: OpenMOSS-Team/MOSS-Image-SFT
• Tags: diffusers, moss-image, controlnet, lora, comfyui, pose, image-to-image

Links

• Model: https://huggingface.co/thedeoxen/MOSS-Image-pose-controlnet
• Files: https://huggingface.co/thedeoxen/MOSS-Image-pose-controlnet/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download thedeoxen/MOSS-Image-pose-controlnet

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