MOSS-Image Control LoRAs in ComfyUI: Canny, Depth, Pose

thedeoxen published seven control LoRAs for MOSS-Image-SFT on November 10: canny, depth, pose, normal, lineart, scribble and segmentation. Each one is a single ~290 MB rank-64 LoRA with its own ComfyUI workflow and control-to-result example grid, so you can pick the map you already have and run it.

They are edit LoRAs, not a new control model. The map goes in as the reference image and the prompt decides what the shapes become.

Depth control (leftmost) with generated results from different prompts

What each LoRA takes

How to run it

1. Load MOSS-Image-SFT: UNet moss_image_sft_int8_convrot.safetensors, CLIP qwen3.5_9b_w4a8_gptq.safetensors with type moss_image, VAE flux2-vae_small.safetensors.
2. Add the LoRA on the UNet at strength 1.0. Drop toward 0.7 if the control is too rigid.
3. Feed the control map as the single reference image into Text Encode MOSS-Image Edit, and set edit_format to ai-toolkit on both the positive and the negative encode. These LoRAs were trained with kv cache on.
4. Sample with 30 steps, CFG 4, euler and simple.

The preprocessors in the workflows (Canny, Depth Anything V2, DWPose, BAE, LineArt, HED, UniFormer) come from the ControlNet Auxiliary Preprocessors pack, not from core ComfyUI. If you already have a map, skip the preprocessor and wire the image straight into the encode node.

Flat-color regions from UniFormer, restyled per prompt

How this differs from the official adapter

The official MOSS-Image color/depth control adapter starts from noise and takes a region map plus a depth map as conditioning. These LoRAs instead treat the map as a reference image, which is why they work with any of the seven map types and why the prompt carries all appearance. The trade-off is that strong controls can hold too tightly: the cards recommend strength 0.7 when the layout looks frozen.

Downloads

Base model: OpenMOSS-Team/MOSS-Image-SFT. Each repo carries its own moss_image_<type>_control.json, which you can drag into ComfyUI once the three base files are in place.