This is a control model trained as an editing LoRA: pass the colored skeleton generated by DWPose (black background, including body, hands, and face) as the sole reference image to the Text Encode MOSS-Image Edit node. The model follows the skeleton pose, while the character, clothing, and scene are determined by the prompt. LoRA linear layer rank 64, convolutional layer rank 16, take step 4500. The author has also released two similar LoRAs for depth map control and flat color block segmentation control. The repository includes ComfyUI workflows and multiple sets of example images.
Why it matters MOSS-Image now supports controllable human poses. The same author has also released two control LoRAs for depth maps and segmentation maps.
Hugs & Likes 11 · Hugs Rising Score 11 · Hugs Downloads 0
License: Commercial use allowed · ComfyUI: Mentioned in model card · VRAM: Not specified
Reasons
- Apache-2.0 license; each set of example images provides the control image and results from three different prompts
Usage
- Load MOSS-Image-SFT, set LoRA strength to 1.0, set the editing format to ai-toolkit (must be set in both positive and negative encoding nodes), 30 steps, CFG 4, euler + simple; extract the skeleton using DWPose from comfyui_controlnet_aux
Risks
- This is not the official MOSS color/depth concatenation adapter; limbs missing from the skeleton cannot be inferred by the prompt
- Trained and recommended only for MOSS-Image-SFT