Request for license clarification
Hello,
We are considering ZhengPeng7/BiRefNet for use in a commercial service. The model would run only on our server, and users would receive only the output images.
The model card indicates MIT, and we found an MIT license in the linked GitHub repository. Could you please clarify whether the Hugging Face model weights are also covered by the MIT License?
If so, could you please share the required copyright notice and consider adding a LICENSE or NOTICE file to this repository that explicitly covers the model weights?
Thank you for your help.
You can find the license I set in the model card of each HF model.
Hi , and thanks for BiRefNet!, I wanted to follow up on this thread because my use is commercial and it is regarding the model weights. I checked the model card as mentioned above and need some clarifications regarding license.
The model card says MIT. Can you confirm that covers the released weights of this checkpoint,The card says it's trained on DIS-TR. The DIS5K terms restrict commercial use of the dataset itself. Is the model fine to use commercial purpose.
Thank you so much..
Yeah, quite a lot of people asked me about it. Let me just paste one of my email responses here for your reference:
You can refer to the related issues in the BiRefNet GitHub repo. I also discussed with claude for quite a lot on this question.
In 3 short sentences: using backbones pretrained on ImageNet is not even 100% fine, since it claims not to be used for commercial use. There are also legal cases that said only the models can recover the original images would conflict with the claim of the dataset authors. Almost all existing open-source models might conflict with this rule in the worst-case scenario.
From my perspective, you can surely use it anyway under the MIT license.
But if you want to be 100% sure, you can also ask about the datasets used for training there, as I listed in the model zoo section in the README.
Hi Peng,
Thank you for your previous clarification regarding commercial use.
We have now traced our exact browser model to:
ZhengPeng7/BiRefNet_lite/model.safetensors
Revision:aa62cd87eafb9cc43056d08ef3615a14628b831d
Our converted browser weights were verified as numerically equivalent to the official checkpoint (378/378 tensors).
We intend to use this model commercially in a browser-based background-removal tool at smarttimetools.com. The model runs entirely on the user's device and user photos are never uploaded.
Could you please confirm that the MIT license on BiRefNet_lite is intended to permit:
- commercial use of this exact model checkpoint; and
- redistribution of a numerically equivalent converted/packed version of these weights for browser inference?
We are separately contacting the DIS5K authors regarding the training-dataset terms.
Thank you very much.
From my perspective, feel free to use it in any way under the MIT license. Good luck with your project.