They also have the resources to build a huge training set, together with people who willingly upload their art and photos to them, which they can use to make the training set better than publicly available data.

One step further, they already have a huge training set. Stock libraries have the luxury of the hard part already being done: labeling. As of today, that's >313M labeled images they can use with no fear of legal woes: https://stock.adobe.com/search/images?filters%5Bcontent_type...

Stable Diffusion was trained on billions of images, of course. But having explored some of LAION-2B, it's clear that Adobe Stock has far better source images and labels.

Just to be really clear what we do and do not train on:

Firefly was trained on Adobe Stock images, openly licensed content and public domain content, where copyright has expired.

https://firefly.adobe.com/faq

We do not train on user content. More info here:

https://helpx.adobe.com/manage-account/using/machine-learnin...

Is it also trained on all of the Adobe Stock images which were created with Generative AIs in the first place? I think that could get into a messy spot too. It looks like there are already millions on there, not all of which is actually labeled as generative.

If you go to https://stock.adobe.com/ and search for `-generative` you can see it filters out ~3M of them, but then go to the "undiscovered content" filter, and it appears to me like the vast majority of it is generative too.

They also know most of their business customers already have GPUs, and often have high-end GPUs, so they're able to tailor solutions to the hardware their customers already have. For example, the speech to text feature in Adobe Premeire runs on local hardware, and is actually pretty good.

Hopefully they'll continue to push the potential for locally run models.