StyleDrop: Text-to-Image Generation in Any Style | Hacker News Reader

StyleDrop: Text-to-Image Generation in Any Style

styledrop.github.io

114 pointsfamouswaffles32 comments

Is this another squandered Google AI showcase? To what purpose does another demonstration serve Google when it is not immediately paired with a product announcement, beta access invitation or source code release? Does it merely make it easier for Google researchers to be poached by other AI companies that are actually releasing AI products? Adobe is actually integrating image generation tools directly in its products. What are you waiting for?

I was thinking the same, recent announcements from Google are research papers leading nowhere. At this point as soon as I see something from Google I know it's either a vaporware or a product that will be killed in next 6-18 months.

> announcements from Google are research papers leading nowhere

And without code or model files, who knows if research is actually reproducible and not cherry-picked to look good?

Attention is all you need had no code or model and still led somewhere.

that's not true, "attention is all you need" presents a very complete model that is replicable.

In what sense it was more complete or replicable than this paper?

It's a research paper, not a product. I don't know why you're against this, but I appreciate researchers that publish papers so knowledge isn't siloed in organizations and contribute to the public good and advancement of technology. Please do more of this!

Straw man, no where did I say that I am against the release of this paper. All I said that was that I don't understand how Google benefits, and I selfishly claimed that more of us could benefit if they created a product. The paper by itself is clearly a gift. I think ultimately it hurts Google to showcase research without product development -- and I also believe the trained model is squandered. If they aren't going to productize it, then they should leak or open source it in a LLaMA fashion. Why the waste? If this was truly a beneficent act somehow despite the weights being locked up in their cloud, where is the carbon footprint data for the training? StabilityAI reports such information for their model releases.

Google doesn't benefit from sharing. Humanity does. It was a beacon for mankind and will continue to be. As for how it can survive on this and how long it can survive on this... well, I guess that's largely depending on how bad sundar drives the company

It's not really sharing tough, is it? No source code, no models, no demo... just a couple pages of text, a couple of vague formulas, insufficient to reproduce results (which by the way should be the whole point of research papers). This is not sharing, this is just showing off.

It's a landing page. The other 99.999% of research papers around the world don't have landing pages for a reason.

Having a project page with extra results is very common in computer vision, even in academia.

I feel like people don't understand what this is, and what google has historically done.

This is research for some math to improve the ability to do a style transfer, and they show it on their own text-to-image generator muse, which they have also published the structure of.

This is what they have typically done, and this is what they didn't do with Bard.

No, they did not release waits for it, you cannot run this on your own computer. But they typically didn't release weights for things like Lambda or Imagen either AFAIK.

This is not a product. This is not a tool for you to use. This is for researchers.

The point of this paper is not to let you run it on your computer. It's to allow other researchers to implement and build on the methods described in the paper.

Well, it’s clearly not a good strategy since it’s what allowed OpenAI & StabilityAI to get all the credit.

The researchers want credit for their work. Google wants to stay ahed of their competitors. Google has three moves:

1) Allow publishing everything including source code => this helps the competitor directly. Bad move.

2) Disallow publishing => the researchers will be tempted to switch jobs for their competitor, since staying at Google will hurt their career. Bad move.

3) Allow publishing, but disallow everything else => this helps the competitors a little, but not too much. The researchers get credit for their work, which removes any incentive they have at switching jobs. Seems like the best compromise.

At least, that's my speculative take on this. Sure, OpenAI & StabilityAI get the credit in the public's eye, but there are also other incentives at play.

I think it's pretty obvious MBA case studies will be based on this in future, just as with Kodak inventing but not pursuing digital cameras, Xerox being responsible for the tech behind multiple billion dollar companies they didn't pursue, etc., etc.

Very interesting point! Hopefully its not the case so the trend can continue...

You’re aware that the peer review process is basically impossible without the release of weights?

Nobody (to good approximation) plays with weights during the peer review process.

Then why make a fancy landing page instead of just linking to the paper?

Anyway, it's not 2000, you can't get away with releasing a paper without code. In the case of AI/ML, you either need to release the weights, or make some web doodad that allows you to use the model. If that's not there, I just assume that the results aren't reproducible.

> Then why make a fancy landing page instead of just linking to the paper?

Because it's worked for the longest time.

Even a year ago, Google was percieved as being so far ahead. These little papers with their landing pages were like sneek peaks into the advanced tech behind the scenes. It's marketing that brought hype to Google's brand and we were all excited for it because none of the big movers felt enough pressure to actually put stuff out so we were all excited about the possibilities.

And we are telling them now that this shtick doesn't work any longer in 2023 post-ChatGPT, post-StableDiffusion, post-LLaMA. Our expectations have clearly shifted. Other companies and organizations are creating actual products and sharing actual models. We are done inhaling vapor.

I don't mean to be rude, and to be clear I wish they did release the weights, but what did they lose here?

You don't approve - so what? Releasing the weights doesn't make them money.

I'm skeptical LLaMa is even useful for facebook commercially at least, they don't make money on it, and I doubt anyone developed brand loyalty, more then likely everyone will use whatever the next, best open model is regardless of who makes it.

Llama and SD still don't come close to midjourney/chatgpt/claude when you look at ease of use and infrastructure cost. These "99% the performance of chatgpt" are laughable if you use them (which I have extensively).

> We are done inhaling vapor.

Okay? What were you about to pay for to begin with here?

EDIT: Just to add, it's not like we got nothing from this, this can likely still be something to try with SD.

Interesting how often "a banana" turned into multiple bananas, and how the coffee machine flip-flops between infusion and grinding, and the towel not only changes number, but also sometimes disappears entirely!

In the last century, machines replaced craftsmen by doing machine-like work (a handmade nightstand vs an industrial nightstand). Now they will also replace the remaining craftsmen, or the last jobs that only a craftsman could do (just look at the example of icons in the suggested style).

But it doesn't matter, soon they will stop doing what we prefer, I guess...

Can we maybe keep comments about general "AI and society" to posts about that are actually about that?

There's about 10 posts on the front page a day about AI and filling each one up with comments about the broader topic just diffuses the debate - both about the post topic in question and about the important issues you're trying to raise.

You're right. For me it was relevant because, for some reason, this is the first time I've experienced this feeling while looking at one of the latest 1000 studies on the subject. However, I must admit that I too am feeling a bit of fatigue when it comes to AI, and I understand the need for specificity.

Super cool, thanks for sharing!

It appears to be fantastic work... work that people can't use.

It seems to be published as research not as a product announcement.

It's confusing that the page's presentation makes it feel like an actual product. The branding of "Style Drop" and the copy makes it feel like something "you" the reader can use right now... and then you can't find a download or sign-up link.

The big question is if Google will be able to capitalize on their research. Or if publishing all those papers just plays into the hands of the competition.

When I want to try Bard, I still get "Bard isn’t currently supported in your country. Stay tuned!". While me and everyone around me already uses ChatGPT, FastGPT, Phind and Perplexity.

Reply on news.ycombinator.com