Licensing question
Hi guys. I was a little confused about the licensing here and wanted to make sure I'm interpreting it correctly.
The CommunityForensics dataset is licensed CC BY 4.0, but CommunityForensics-Small is licensed CC BY NC SA 4.0. The models you trained are licensed MIT.
Am I correct in understanding that I could train my own model on the full CC BY 4.0 CommunityForensics dataset, use that model commercially, and release the resulting weights under MIT (assuming the appropriate attribution for the dataset)?
Separately, is the -Small dataset intentionally restricted to non-commercial model training, or is the NC license only intended to restrict redistribution/use of the dataset itself?
I also noticed the dataset contains outputs from generators with their own licenses/TOS. Are there any upstream licensing restrictions you're aware of that would prevent commercial use of a model trained on the full dataset?
Just want to clarify before I build anything on top of it and accidentally step on anyone's toes. Thanks!
Hi trentmkelly,
The reason why CommunityForensics-Small is licensed CC BY NC SA 4.0 is because it contains the 'real' data as well for research convenience. Some of those 'real' datasets do not allow commercial use (e.g., CelebA, FFHQ). The CommunityForensics dataset only contains the generated images, so I licensed it as CC BY 4.0, which allows commercial use (with appropriate citation/attribution). In other words, commercial use of the CommunityForensics dataset is OK with me with appropriate attribution.
The use of the -Small dataset is limited by the real dataset's terms for non-commercial research purposes. However, you are free to use other sources of 'real' data.
Vast majority of the models are licensed using the CreativeML Open RAIL-M license. The full list can be found on our paper, page 16 (last page), Figure 10. To my knowledge, the use of the model outputs to train an AI-generated image detector should be permitted under transformative use, but you may want to consult your lawyer on whether it would be acceptable for your use case.
Hope this helps, Thanks.
Thank you very much!