In LooseControl, the authors trained a LoRA of ControlNet-depth, but now few libraries or frameworks support LoRA of ControlNet, so they hacked through ControlNetModel of diffusers with UNet2DConditionLoadersMixin.

However, we can't run the code in frameworks like A1111's WebUI or ComfyUI, so we fused the weights of ControlNet-depth and LooseControl to make it work in any frameworks. For details, please refer to the script.

Important Note: The authors of LooseControl did more than just training a LoRA. Let's not forget that. Please refer to the original paper and code for more details.

Download the fused ControlNet weights from huggingface and used it anywhere (e.g. A1111's WebUI or ComfyUI) you can use ControlNet-depth to loosely control image generation using depth images.

Example folder contains an simple workflow for using LooseControlNet in ComfyUI.

If you like it, you can contribute by:

• Upvote this issue in diffusers repo or possibly make a PR to resolve it.
• Bring consistency mechanisms devised in LooseControl to frameworks like A1111's WebUI or ComfyUI.
• Bring box editors to frameworks like A1111's WebUI or ComfyUI.
• Perhaps train a better LooseControlNet

The extra code we add is released under MIT License and the fused weights are released under Apache 2.0 License, which follows the original license, MIT License, of LooseControl and Apache 2.0 License of ControlNet.

This is the official repository for LooseControl:

LooseControl: Lifting ControlNet for Generalized Depth Conditioning

Shariq Farooq Bhat, Niloy J. Mitra, Peter Wonka

[Project Page] [Paper] [Demo 🤗] [Weights (3D Box Control)]

@misc{bhat2023loosecontrol,
      title={LooseControl: Lifting ControlNet for Generalized Depth Conditioning}, 
      author={Shariq Farooq Bhat and Niloy J. Mitra and Peter Wonka},
      year={2023},
      eprint={2312.03079},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Please refer to its official repository for more details.