ONNX Models & Logik Users

[allklier]

As we’re looking to make more ONNX models available to the group, we should discuss the legal caveats we should observe.

There’s a good reason why it took a while for ADSK to come out with new ML tools. The latest thinking by all these companies (Flame, Nuke, Baselight) is that they make the inference nodes available for you to run models, but they don’t touch the models other than their own proprietary ones.

The whole reason is the question of legal liability of the training data. That liability lies with the person using the model, not with the person making the inference node. They’ve put the burden on us to deal with the legal issues. Which is totally fair. But also complex.

Ultimately the liability with any of these ONNX models other than the MLTW that is bundled with Flame 2025.1 is with any one of you who runs one of these models on a job, gives the render made with it to your client and collects a check. The pattern of the last decade across most of life - push more of the burden downstream on the little people.

But the specific question I have - as we’re putting links to ONNX models on this public forum and more importantly on the Logik Portal, what type of legal language do we need to include to not get caught up in the cross fire. It puts us into the chain of distributing materials of at times questionable licensing status.

For reference Foundry includes this language on their Cattery page:

Foundry does not certify or approve, nor are we responsible for, the licensing or legal framework to the underlying pre-trained model, associated training data, and dependencies. The User should undertake independent investigation of the underlying files in so far as it is relevant to the User’s intended use.

Do we need to put a variation of this into the Logik Portal page (@MikeV) and also into any regular Logik post that has links to ONNX models to download (like I included a Dropbox link yesterday)?

And on that account - should the Logik Portal identify the specific contributor of the ONNX model, or should we keep that as a community resource to not put a singular person into the line of fire?

Thoughts @andymilkis @randy?

[fredwarren]

@allklier

The whole reason is the question of legal liability of the training data. That liability lies with the person using the model, not with the person making the inference node. They’ve put the burden on us to deal with the legal issues. Which is totally fair. But also complex.

I just want to add that one of the main reason is that many models are open source but cannot be distributed for commercial purposes. I think this is why the cattery models are available for free from a webpage rather than being bundled with Nuke.

[allklier]

Thanks for that extra point @fredwarren.

It reinforces the question. Since 99% of what all of us do is work for commercial clients rather than art for the sake of art, us (individuals and Logik) distributing variations of these models for other people to use them for commercial purposes puts us into that exact same gray space.

Unlike ADSK, we don’t get paid for what we do here on the forum, so it’s a different shade of gray. But it’s not white. We’re not commercially benefiting from the distribution, but we need to maintain the legal caveats that these are only used within each model’s licensing terms (and in proxy, the licensing terms of its training data to the extent that they can be ascertained).

[randy]

I don’t have a legal position to offer because Andy and I don’t own nor operate the Logik Portal. We have nothing to do with it. Flame Logik LLC owns and administrates the Logik Forums, the Logik Discord, and Logik Academy Pro.

Other community services like the Logik Portal, and the Logik Matchbox Shaders at Logik-matchbook.org are owned and operated by someone else.

[allklier]

Fair enough. Totally understand from a business of Logik perspective.

Leaves each community member to decide on their own path.

For now I’ve added a disclaimer to my post, and will hold-off on submitting models to Logik Portal until I can see where we land with this.

Happy to contribute with cautious optimism.

[randy]

[MikeV]

Good points @allklier. I just put out a new version of the Logik Portal that hopefully address some of this.

The names of people who’ve uploaded Inference node has been removed. If people want to take credit for uploading they’re more than welcome to add their name into the description of the node. Anyone who has already contributed can email me and I’ll be happy to add their names into the descriptions of the nodes they’ve already uploaded.

Any nodes that get uploaded are going to need to be open source and come with a link so that people can read up on what the licenses are for any node and the source of the material the node was trained on.

Finally when downloading a Inference node there will be a disclaimer shown warning users to read up on the nodes before using them.

For versions 6.0.0 and 6.1.0 of the Logik Portal I’ve removed the ability to download the Inference nodes. The Inference tab will show up empty. Anyone wanting access to them again will have to update to 6.2.0 through either the Portal or here: Logik Portal — Flame Python Scripts.

Hopefully this will alleviate some concerns with uploading new nodes and give fair warning to anyone downloading them.

Mike

[motoronlysync]

@allklier

…Unlike ADSK, we don’t get paid for what we do here on the forum…

Wait, I am supposed to get paid for this? Randy, did my check get lost in the mail again?

[allklier]

Ah the precision in language bites again….

This was in context of ML model creation/conversion and whether our activities could be seen as commercial activity.

All general forum contributions in terms of time and effort are voluntary, and very much appreciated in particular for our ADSK friends, going above the general norm.

[randy]