MOSS-Image-SFT gets Apache 2.0 ControlNet adapters for pose, depth and edges

It is a familiar frustration for anyone who generates images with AI: the prompt describes “what,” but offers little control over “where.” You ask for a person jumping, and the model shows them standing. You request a specific camera angle and get a random one. That is what ControlNet was designed to address: alongside the text, you supply structural guidance, and the model builds the image around it.

MOSS-Image-SFT now has such a toolkit. It has been released under the Apache 2.0 license, meaning it can be used in commercial projects without separate approval.

What has been released

The lineup includes seven adapters: OpenPose/DWPose (body skeleton and keypoints), Depth (depth map), Canny (edges), Lineart (line drawing), Normal Map (surface normals), Scribble (rough sketches) and Segmentation (region labels).

They all follow the same principle. Each module is trained as an editing adapter: the input image defines the structure of the frame, while the text specifies everything else — the character’s appearance, materials, surroundings and style.

Here is a simple example. Take a photo of someone in the pose you want, run it through DWPose and extract a skeleton. Then enter a prompt: “an astronaut in a rusty spacesuit against a desert backdrop.” The pose comes from the source image, while the model generates everything else. The same technique with Depth preserves a room’s perspective while completely changing its interior, and with Canny it preserves the silhouettes of objects.

Technical details

The adapters use LoRA: rank 64 on linear layers and rank 16 on convolutional blocks. Training was performed with the KV cache enabled. For practitioners, another detail matters more: compatibility with quantized weights is claimed. The UNet runs in INT8, while the Qwen3.5-9B text encoder uses W4A8 GPTQ. This configuration runs on consumer graphics cards with 16–24 GB of VRAM.

Ready-made workflows are available for ComfyUI. They use the Text Encode MOSS-Image Edit node in ai-toolkit mode. There is no need to build the chain from scratch: load the workflow, plug in the weights and supply a reference image.

Weights for OpenPose, Depth and Canny have already been published on Hugging Face. Judging by the links, they were uploaded by a third-party contributor (thedeoxen), rather than the OpenMOSS team responsible for the base model.

Limitations and how to approach them

The post includes neither benchmarks nor comparisons with ControlNet implementations for other models. It is therefore too early to judge the quality and reliability of the structural control: you will need to test for yourself how accurately poses hold up in complex scenes and whether edges drift when the style changes. In addition, weight links are available for only three of the seven adapters. The post does not make clear where to find Lineart, Normal, Scribble and Segmentation.

The practical takeaways are straightforward:

• If you already run MOSS-Image-SFT in ComfyUI, try Pose, Depth and Canny and compare them with your current pipeline using your own reference images.
• If you are choosing an open model for a product pipeline (character generation, storyboards or interiors), structural control under Apache 2.0 is a significant advantage. It removes the need to rely on closed APIs for predictable composition.
• If you have a 16 GB card, start with the quantized configuration and check how INT8 and W4A8 affect detail.

Ultimately, this is not an architectural breakthrough, but a missing piece of infrastructure. Yet pieces like this are precisely what turn an open model from a demo into a practical tool.