When Adobe Inc. released its Firefly image-generating software last year, the company said the artificial intelligence model was trained mainly on Adobe Stock, its database of hundreds of millions of licensed images. Firefly, Adobe said, was a “commercially safe” alternative to competitors like Midjourney, which learned by scraping pictures from across the internet.
But behind the scenes, Adobe also was relying in part on AI-generated content to train Firefly, including from those same AI rivals. In numerous presentations and public postsabout how Firefly is safer than the competition due to its training data, Adobe never made clear that its model actually used images from some of these same competitors.
why would they do this, doesn’t that reduce the quality of training dataset?
Depends how it’s done.
Full generative images would definitely start creating a copying error type problem.
However it’s not quite that simple. An AI system can be used to distort an image. The derivatives force the learning AI to notice different things. This can vastly extend the pool of data to learn from, and so improve the end AI.
Adobe obviously decided that the copying errors were worth the extended datasets.
Supplementary synthetic data increases the quality of the model.
Correct. To a certain extend one can add AI data into AI, too much and you add noise, making the result worse, like a copy of a copy.
Yes, though that’s not what they’re doing. They train on images uploaded to their marketplace and, of course, some of these are AI generated.
It’s fine as long as it’s not the majority.
It doesn’t really matter how much it is. An image is an image.
Data augmentation is a thing since a long time, but of course if the majority of your data is synthetic your model will suck on real world data. Though as these generative models get better and better at mimicking real world data and we select the results we want to use (removing the nonsense and hallucinations, artifacts etc.), we’re still feeding them “more data”.
I guess we’ll have to wait and see what effect it’ll produce on future models. I think overall the improvements on LLMs have been good, even at slow steps we’re still figuring out how to better turn them into useful tools. I don’t know how well the image generation models have improved in the last 2 years though.
Yes, that’s one way of putting it. What gets into the Adobe stock database is already curated. They also have the sales and tracking data.
Also yes on this. It doesn’t matter if your data is synthetic but only if it’s fit for purpose. That’s especially true in this case, where the distinction between synthetic and real is so unclear. You’re already including drawings, renders, photomanips, etc. I have no idea what kind of misconception people have that they would think it matters if some piece of digital art is AI generated.
I’m just talking about synthetic images affect model quality.
It doesn’t matter how the image was made. It only matters what it is like and how it is used to affect the model.
That’s what I’m saying. Synthetic images can help your model look better, but if you’re aiming for “realistic” output, but synthetic images are fundamentally not real images and too many will bias your model in a slightly different direction.
No.
I feel I should explain this but I got nothing. An image is an image. Whether it’s good or bad is a matter of personal preference.
I’m not so sure about that… if you train an ai on images with disfigured anatomy which it thinks is the “right” way it will generate new images with messed up anatomy. It gives a feedback loop, like when a mic picks up its own signal.
Well, you wouldn’t train on images that you consider bad, or rather you’d use them as examples for what not to do.
Yes, you have to be careful when training a model on its own output. It already has a tendency to produce that, so it’s easy to “overshoot”, so to say. But it’s not a problem in principle. It’s also not what’s happening here. Adobe doesn’t use the same model as Midjourney.
Midjourney doesn’t generate disfigured anatomy. You’re think of Stable Diffusion which is a smaller model that can generate an image in 30 seconds on my laptop GPU. Even SD is pretty good at avoiding that, with decent hardware and larger models (that need more memory).
When you process an image through the same pipeline multiple times, artifacts will appear and become amplified.
What’s happening here is just nothing like that. There is no amplifier. Images aren’t run through a pipeline.
The process of training is itself a pipeline
Yes, but the model is the end of that pipeline. The image is not supposed to come out again. A model can “memorize” an image, but then you wouldn’t necessarily expect an amplification of artifacts. Image generators are not supposed to d lossy compression, though the tech could be used for that.
If the image has errors that are hard to spot by the human eye and the model gets trained on these images, thoses errors that came about naturally on real data get amplified.
Its not a model killer but it is something to watch out for.
Yes, if you want realism. But that’s just one of the things that people look for. Personal preference.
Invisible artifacts still cause result retardation, realistic or not. Like issue with fingers, shadows, eyes, colors etc.
deleted by creator