Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Imagen: An AI system that creates photorealistic images from input text

imagen.research.google

Post reply on HN
Print view View original

71–80 of 233 posts

Re: Imagen: An AI system that creates photorealistic images from input text

#71

Post by glenneroo »

hackernewds wrote: post #40
NegatioN wrote: post #33

Earlier quoted context omitted.

The previous models were either 1. Limited in their capacity to create something that looked very cool, or 2. Gigantic models that needed clusters of GPUs and lots of infrastructure to generate a single image. One major thing that happened recently (2ish weeks ago) was the release of an algorithm (with weights) called stable diffusion, which runs on consumer grade hardware and requires about 8GB of GPU RAM to generat…

Is Lexica finding results previously computed? Or generating them? I could only work with very simple queries like "photo of a cat".

It's ~1.5 million entries inputted by users during the beta period on Discord.

Re: Imagen: An AI system that creates photorealistic images from input text

#72

Post by kyriakos »

culi wrote: post #18
bottlepalm wrote: post #13

Earlier quoted context omitted.

Ok I’m just going to list ideas, feel free to add more * Text to 3d model * Text to video clip * Illustrations for newspapers * Generating pictures for food menus * Generating a music video from a song * Generating pictures or even an entire movie for a book * Interior design ideas * Product design ideas I have noticed that leveraging these tools isn’t easy. It requires a fair bit of creativity to come up with the pr…

Someone on the midjourney discord is making a comic with images generated by the bot. Another person is making a Magic The Gathering card pack with generated art (it looks good too!). People are already using it for stock photos One murky area we're still far away from but I'm curious to follow the developments on: AI-generated movies. Once generated clips gets good, what if some movie buff can just generate a movie,…

How do they get consistent looks? I mean if you are using it for a comic book your characters should look the same throughout the story. From my experience, the results between prompts are completely different even with minor variations.

Re: Imagen: An AI system that creates photorealistic images from input text

#73

Post by glenneroo »

schappim wrote: post #30
panabee wrote: post #16

past threads: * https://news.ycombinator.com/item?id=31484562 * https://news.ycombinator.com/item?id=32589185

Is this being reported as "new news" because it is available to the public?

AFAIK nothing is released yet? They are probably still trying to, like Dall-E 1/2, remove every every image and ban every word that might generate even the slightest hint of controversial imagery before releasing it to the public. To be fair, Stable Diffusion spent months basically doing the same on Discord with thousands of beta users, and had an army of moderators flagging images which were subsequently removed in order to prevent people from generating porn. Didn't help though, as people got access to the beta models (for "research" purposes) a few weeks before the official release and are using that to e.g. create "porn"[0].

[0] pornpen.ai: https://news.ycombinator.com/item?id=32572770

Re: Imagen: An AI system that creates photorealistic images from input text

#74

Post by Roark66 »

Every single time I see an article about a new AI model that has a section called "societal impact" I know immediately they are not releasing the model, the training set, nothing...

It seems to be the kind of bullshit statement that those companies put in place of "we paid $500k training this model and we're not giving it for free to anyone".

Re: Imagen: An AI system that creates photorealistic images from input text

#75

Post by johnthewise »

PoignardAzur wrote: post #64
MathYouF wrote: post #21

Earlier quoted context omitted.

Like anyone deeply in a field I know maybe several thousand people who could probably give a better answer, but I figure I'll give an effort to provide one since I don't see any good ones posted yet. The moment everyone knew this was going to be big was in 2019 when StyleGAN came out. They used a lot of tricks like aligning face features (like eyes) and had all their pictures of a single domain (the most famous being…

Interesting! I knew about transformers, CLIP and diffusion, but pixel patch encodings are new to me. Can you give me more details / point me towards an explainer? A quick duckduckgo search didn't help.

I don't quite remember whether it was first used in Vit paper[1], but it's a fairly straight forward idea. You take the patches of an image like they are words in a sentence, reduce the size of the patch(num_of_pixel x num_of_pixel) with a linear projection so that we can actually process it and get rid of sparse pixel information, add in positional encodings to put in location information of the patch and treat them as how you treated words in language models from that point on with transformers. Essentially, words are human constructed but information dense representation of language but images do have quite sparsity in them because individual pixel values don't really change much of an image.

1: https://arxiv.org/pdf/2010.11929.pdf

Re: Imagen: An AI system that creates photorealistic images from input text

#76

Post by glenneroo »

culi wrote: post #23
8n4vidtmkvmk wrote: post #20

Earlier quoted context omitted.

those were generated? those look dope

Generated and paying for MidJourney gives you the copyright to them so you can use them for whatever projects you want

Which feels a bit sketchy(?) to me, seeing as all models are built on imagery scraped from the net without anyone's permission. It's one reason why I've spent the last month training my own models using my own imagery. If these stock photo sites had any brains, they would also start training models on images in their databases, especially since they already have everything sorted into categories based on keywords (which I'll spend the next year doing, until I can get img2text tools working in recursive batch mode).

Re: Imagen: An AI system that creates photorealistic images from input text

#77

Post by ansible »

polotics wrote: post #3

Not a single human face in these samples. I wonder how well it does on faces.

The samples I was shown definitely had a theme: robots, raccoons, hats, and raccoons wearing hats.

Our brains are very thoroughly wired to detect faces in general, and flaws in faces. So we have a very, very high standard for what passes muster.

We are apparently much more forgiving with regards to what raccoons and corgis look like.

Re: Imagen: An AI system that creates photorealistic images from input text

#79

Post by FairlyInvolved »

echelon wrote: post #57
mshenfield wrote: post #28

These tools are amazing for prototyping. I had an idea for a promotional poster, and seeing my idea just by writing it felt like magic. The generated image had too many artifacts to use, but gave me a guideline to follow when creating the real thing in Pixlr. AI content generation (text, image, source code, video, music) will be a huge boon for prototyping where applied judiciously.

Google hasn't released squat . Google's product is vaporware and we shouldn't afford them any airtime until they release something usable. They're just trying to butt in and get press off of the backs of the teams actually working in the open, and that's super lame. Release your model, Google, or stop bragging and talking over the others here. You're greedily sucking oxygen out of the conversation, and as a trillion…

I kind of get the sentiment about openness but I think it's way more nuanced than you are making out.

There are very good reasons for withholding SOTA models, primarily from the info hazard angle and avoiding escalating the capabilities race which is basically the biggest risk we have right now.

Google / Deepmind have actually made some good decisions to try and slow down the race (such as waiting to publish).

Post reply on HN