Encoder-Based Domain Tuning for Fast Personalization of Text-to-Image Models | Hacker News Reader

Encoder-Based Domain Tuning for Fast Personalization of Text-to-Image Models

tuning-encoder.github.io

65 pointsminimaxir15 comments

Kind of ghoulish to decide her art is fair game to rip off scarcely three years after her death... and that it's so acceptable it should be your system's example.

https://mashable.com/article/qinni-qing-han-artist-died

What is wrong with using the style of an entity which by definition of being dead is beyond caring? Should we hold off in perpetuity, such that we not copy the style of Michaelangelo lest sensibilities be ruffled? What even is this platonic ideal form of one's "style"? Is this interpretation of the artist's work not a creative endeavor in itself? It didn't create itself de novo, after all. A human ultimately directed a machine to perform a task. If you want to really get down to how I feel about copyright in the age of AI, consider that a number can't be copyrighted in the US and yet any image file can ultimately be represented by a sufficiently large binary number. The creative spirit of the American people has been bottled up by corporations for far too long.

Her sister asked you stopped. https://twitter.com/QinniArt/status/1591167678684999681 https://twitter.com/QinniArt/status/1591176309094649856

It's okay for humans to riff off this style, though? Because I continue to see fleshy artists doing it.

I keep seeing artist use acrylic paint, such a blatant rip off of long established art styles. This has to stop.

EDIT: let me take this opportunity to also point out how better copyright enforcement could have saved us from the likes of this fraud Picasso, who just ripped off the style of Braque

I'd never heard of her, but her art is nice. The authors credit her, as they should, and that should be enough as we're only talking about inspiration here.

It's hardly ghoulish to make new works in the style of a late artist, when you're open about what you're doing and giving credit.

Using the word "entity" instead of "person" is a real, uh, choice here. Invocations of Michelangelo and machines creating themselves are not argued in good faith against anything I actually said.

Ultimately, no, I don't want to really get down to how you feel about copyright in the age of AI. That is a total distraction from how, completely independently from whether it's illegal, deciding your demo should be "hey let's use computers to replace a widely beloved artist who died too young and whose surviving family doesn't want us to" is wildly heartless.

Responding as though she was a corporation, as though I'd been talking about IP law, as though that's the only lens worth taking on this – sort of underlines the ghoulishness.

This is like Dreambooth, but takes 5s instead of 5 minutes, by doing fewer larger fine-tuning steps.

Keep in mind that TI and LORA are already much faster than dreambooth (but admitedly not 5s fast).

LEAP trains custom embeds in even less time (a couple of minutes)

https://github.com/peterwilli/sd-leap-booster

"a couple of minutes" would be much slower than the new method presented in this article?

You should add more example that aren't famous people. In the "Comparisons to Prior Work" image you can see that even pure SD generated recognizable results.

So the encoder takes one image as input and outputs a bunch of words describing it? Or there is some kind of word embedding that's separate from the image embedding?

You can encode an image just like you encode a text prompt. The diffusers library does this in the image variation pipeline: https://github.com/huggingface/diffusers/blob/main/src/diffu...

And you can merge that encoded image with an encoded text prompt, if you wish.

This is presumably something different if its actually altering the weights of the model. It sounds like they figured out how to run a huge learning rate without training exploding?

very cool, will need to test :)

Reply on news.ycombinator.com