Live data from Hacker News

Adobe Firefly: AI Art Generator

adobe.com

301–310 of 496 posts

Re: Adobe Firefly: AI Art Generator

#301
post #297

Earlier quoted context omitted.

I've looked a few times, but have not seen any research on assigning provenance to the weights used in a particular inference run. It's a super interesting space for a bunch of reasons. But the naive approach of having a table of how much each individual training item influenced every weight in the model seems impossibly big. For DALL-E 2's 6.5B parameters and 650m training items, that's 4.2 quadrillion associations.…

https://arxiv.org/abs/1703.04730 > How can we explain the predictions of a black-box model? In this paper, we use influence functions -- a classic technique from robust statistics -- to trace a model's prediction through the learning algorithm and back to its training data, thereby identifying training points most responsible for a given prediction.

Oh thank you! Will go read and digest.

Re: Adobe Firefly: AI Art Generator

#302

Adobe has lost any good sentiment I had for them with their forced subscriptions and dark patterns. I use their products begrudgingly, it’s a sunk cost fallacy from decades of muscle memory. My reaction to seeing any announcement from them is, yeah whatever Adobe.

Agreed. Adobe has lost my complete trust and I've learned to avoid their tools. I honestly don't know what they can do to win me back. They're a damaged brand.

Re: Adobe Firefly: AI Art Generator

#303

Adobe has lost any good sentiment I had for them with their forced subscriptions and dark patterns. I use their products begrudgingly, it’s a sunk cost fallacy from decades of muscle memory. My reaction to seeing any announcement from them is, yeah whatever Adobe.

Once you get stranglehold of the market place, you can pretty much do whatever you want. Adobe... Microsoft... Apple... Google... all pulling the same lever.

Re: Adobe Firefly: AI Art Generator

#304
post #139

Earlier quoted context omitted.

One of the great aspects of open-source stable diffusion (civitai.com et al.) is there's a model for every purpose. Does your inpainting model work with every style? Or is it going to have trouble matching the content for e.g. specific fanart?

It would have trouble matching on trademarked styles, or individual artists / creators styles. One of the primary goals for Firefly is to provide a model that can generate output that respects creators and is safe for commercial use. (I work for Adobe)

I don't really understand the negative comments on this. Though a hacker by heart, I'm a designer first and foremost

And I'm extremely eager to get my hands on AI tools that let me extend my capacities based on _my own_ styles & context, and that is focused enough on this scope to evade future legal obstacles when used in production

Very excited to try this tool!

Re: Adobe Firefly: AI Art Generator

#305

Earlier quoted context omitted.

I am definitely glad to see attention being paid to ethical sourcing of training images but I am curious: did the people who made all those stock images get paid for their work being used for training? Did they check a box that explicitly said "Adobe can train an AI on my work"? Or is there a little clause lurking in the Adobe Stock agreement that says this can be done without even a single purchase happening?

I'm not a lawyer and I don't work for Adobe. :) The contributor agreement linked from here[1] is this: [2] "You grant us a non-exclusive, worldwide, perpetual, fully-paid, and royalty-free license to use, reproduce, publicly display, publicly perform, distribute, index, translate, and modify the Work for the purposes of operating the Website; presenting, distributing, marketing, promoting, and licensing the Work to u…

The terms seem like legalese for "you pay me money now and get to do anything with it". It doesn't seem far-fetched for training AI models to be a valid use-case. This is way better than scraping the whole internet, for art by artists who have had no commercial arrangement with Adobe.

Re: Adobe Firefly: AI Art Generator

#308
post #11

What this reinforces is that unlike with previous big innovations (cloud, iphone, etc), incumbents will not sit on their laurels with the AI wave. They are aggressively integrating it into their products which (1) provides a relatively cheap step function upgrade and (2) keeps the barrier high for startups to use AI as their wedge. I attribute the speed at which incumbents are integrating AI into their products to a…

Except for Chat GPT I havent yet seen super impressive implementations. Dall-e, codepilots, text to speech etc. are still all not good enough to use for more then playing around.

However this landingpage looks amazing.

Any good other tips?

Re: Adobe Firefly: AI Art Generator

#309
post #117

Earlier quoted context omitted.

I can train a LORA on my own PC in less than an hour. Good luck opting out of that.

What does that matter? Generate as much as you want for your own personal reasons. It's about actually being able to use that content legally (and commercially) that matters to most in this conversation.

AI training is a one-way operation - you can't reconstruct the dataset from a model/lora/ti. Unless it's something really blatant like real people, widely recognised copyrighted characters like Batman or Iron Man - it's going to be hard to prove that someone used your art to train an AI model. I'm not required to publish my model or the datasets that I used anywhere.

Re: Adobe Firefly: AI Art Generator

#310

Earlier quoted context omitted.

I am definitely glad to see attention being paid to ethical sourcing of training images but I am curious: did the people who made all those stock images get paid for their work being used for training? Did they check a box that explicitly said "Adobe can train an AI on my work"? Or is there a little clause lurking in the Adobe Stock agreement that says this can be done without even a single purchase happening?

Nobody owes creators who have been paid fully for their work "extra" compensation just because AI is involved. Assuming they have been paid, the work belongs to Adobe.

Defining how “fully” paid a creator has been is the entire point of license agreements. It defines the extent of how much the rights have been purchased away from them.

It merits investigation as to have these creators been “fully” paid to the extent that they have no claim to any future royalties and can have no objection to their work being used as training data.

Post reply on HN