InstantX/InstantID · Unfortunately, this is non commercial usage only and cannot be Apache 2.0 if it is using Insight Face

Jan 25

Just like the Face ID IP adapter, this uses Insight Face, which makes it not usable for commercial purposes.

"The training data containing the annotation (and the models trained with these data) are available for non-commercial research purposes only."

I hate non commercial licenses, and truly appreciate the attempt to release it under a more permissive license, but it unfortunately cannot be if it relies on insight face embeddings.

wanghaofan

InstantX org Jan 29

We agree, we plan to train on other face encoders that support commercial license.

olivert

Feb 6

Shouldn't they be able to release their own code and model under Apache 2 as long as they don't bundle the Insight Face encoder with their code (which they don't) and not release the Insight Face assets under anything else than their non-commercial license? (which they also don't)

This is what they did. They strictly differentiate between the license of their own work (which is released permissively for everyones benefit) and between the Insightsface antelopev2 face encoder which they have no control over and they refer people to Insightface when asked about the license of the face encoder, which is not bundled with their code but has to be downloaded separately.

ostris

Feb 7

@olivert Unfortunately, that is not how it works. Anything trained on the insight face embeddings has to inherit the license. A model trained on a dataset, generated with a model trained on a dataset that has a single image generated by a model trained on a dataset with a single image pulled from a non-com dataset in it, inherits the non-commercial license. It sucks and I wish everyone would completly ignore EVERYTHING released with a non commercial license and forbid it from being on any platform, because it is literally a license virus and poisons absolutely everything it touches.

The code can be released under any license, but the weights, if trained from Insight Face embeddings, cannot. If It could, I would use student / teacher to train a MIT license insight face clone tomorrow and solve all of this, as their code is all MIT, but it just doesn't work that way.

olivert

Feb 7

•

edited Feb 7

I get what your saying and it is true that the way you described the propagation of license restrictions seems to become common practice in the community. But I wonder if that kind of license restrictions actually can legally propagate through model weights or if people just adopted this practice blindly due to them being overly cautious missing to acknowledge that there is a big difference between model weights and source code.

Just because somebody puts some limitations on something doesn't mean it is legal or enforceable.

There are a lot of recent examples from Licenseholders overreaching and trying to put non-enforeable limitations on users. E.g. just few months ago Insight Face got known for trying to get Youtube videos deleted that talk and showcase their work arguing their non-commercial license doesn't allow such showcases since Youtubechannels make money from ads. That is ridiculous and non-enforceable as that would make them the only company in the world able to dictate how, when and by whom their products are allowed to be reviewed and talked about. No company has that kind of censorship power... didn't stop them from trying (and losing).

Also Artists would like to prohibit weights being trained on their work, but they cannot. (as the recent lawsuit against stabilityAI made clear). Artists however would be able to prohibit their work against being included in an apache 2 repo on github. Model weights differ from the way a license can be enforced on them vs how it can be enforced on source code. Common practices that people try to adopt from the history of open source code, does not necessarily apply to model weights.

If Artists could prevent the creation of model weights by putting some license restriction on their Pictures, they would not even have to resort to arguing over fair use vs copyright and just put some ("don't use my pics for training") license restriction and be done with it.

Insight face might be able to limit how their face encoder gets distributed, the same way an artist can limit how his artwork gets distributed. But Seeing how Artists cannot limit how NeuralNets get trained with their artwork I question whether insight face similarly cannot limit how NeuralNets get trained from numbers their face encoder outputs.

Related to your student / teacher clone example: I know it sounds surprising, but that could actually work and not much could be done against that. OpenAI has put license restrictions on their API outputs to prevent exactly such trainings and there is a lot of speculation that this is not enforceable at all, they know it is not enforceable and they just put it there for deterrence. Most what OpenAI could do is ban users/companies from further usage of their APIs citing breach of TOS but that is not the same as being able to prohibit the training and sue the other party to delete their weights.

If we think about it... if it would be that easy why isn't the New York times and every other news outlet just putting a non-commercial license or "no-training" license on their work to prevent OpenAI from using their articles as training data instead of having to resort to suing over the more-difficult-to-prove copyright infringement and having to argue against fair use as an additional risk?

How can OpenAI say on one hand training on articles from news outlets is "fair use" and doesn't require permission but then turn around and say the text that their APIs are outputting cannot be used for training other models and DO require OpenAIs permission before being used as such? They know the restriction is not enforceable but putting it in might deter most from trying (at least in the short term).