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

DALL·E now available in beta

openai.com

Post reply on HN
Print view View original

341–350 of 579 posts

Re: DALL·E now available in beta

#341

Post by Ajedi32 »

TigeriusKirk wrote: post #252
dave_sullivan wrote: post #248

Earlier quoted context omitted.

I think people don't realize how huge these models really are. When they're free, it's pretty cool. But charge an amount where there's actual profit in the product? Suddenly seems very expensive and not economically viable for a lot of use cases. We are still in the "you need a supercomputer" phase of these models for now. Something like DALLE mini is much more accessible but the results aren't good enough. Early ear…

What are the resources at work here? What are the resources needed to train this model? If someone just gave you the model for free, what resources would you need to use it to generate new results?

Training is obviously very expensive, and ideally they'd want to recoup that investment. But I'm curious as to what the marginal cost is to run the model after it's trained. Is it close to 30 images per dollar, like what they're charging now? Or do training costs make up the majority of that price?

Re: DALL·E now available in beta

#342

Post by wishfish »

xnx wrote: post #76

I fully expect stock image sites to be swamped by DALL-E generated images that match popular terms (e.g. "business person shaking hands"). Generate the image for $0.15. Sell it for $1.00.

Makes me imagine stock image sites in the near future. Where your search term ("man looks angrily at a desktop computer") gets a generated image in addition to the usual list of stock photos.

Maybe it would be cheaper. I imagine it would one day. And maybe it would have a more liberal usage license.

At any rate, I look forward to this. And I look forward to the inevitable debates over which is better: AI generation or photographer.

Re: DALL·E now available in beta

#343

Post by zionic »

isoprophlex wrote: post #296

Wait until someone trains a model like this, for porn. There seems to be a post-DALLE obscenity detector on openAI's tool, as so far I've found it to be entirely robust against deliberate typos designed to avoid simple 'bad word lists'. Ask it for a "pruple violon" and you get purple violins... you get the deal. "Metastable" prompts that may or may not generate obscene (content with nudity, guns, violence as I've fou…

Honestly that part pisses me off. Who cares if their AI "makes porn" or something "offensive".

Re: DALL·E now available in beta

#344

Post by visarga »

harpersealtako wrote: post #299
napier wrote: post #282

Earlier quoted context omitted.

GPT3 has seen similar lobotomization since its initial closed beta. Current davinci outputs tend to be quite reserved and bland, whereas when I first had the fortunate opportunity to experience playing with it in mid 2020, if often felt like tapping into a friendly genius with access to unlimited pattern recognition and boundless knowledge.

I've absolutely noticed that. I used to pay for GPT-3 access through AI Dungeon back in 2020, before it got censored and run into the ground. In the AI fiction community we call that "Summer Dragon" ("Dragon" was the name of the AI dungeon model that used 175B GPT-3), and we consider it the gold standard of creativity and knowledge that hasn't been matched yet even 2 years later. It had this brilliant quality to it w…

So it got worse for creative writing, but it got much better at solving few-shot tasks. You can do information extraction from various documents with it, for example.

Re: DALL·E now available in beta

#345

Post by wongarsu »

jnovek wrote: post #316
peteforde wrote: post #228

I have been having a blast with DALL-E, spending about an hour a day trying out wild combinations and cracking my friends up. I cannot imagine getting bored of it; it's like getting bored with visual stimulus, or art in general. In fact, I've been glad to have a 50/day limit, because it helps me contain my hyperfocus instincts. The information about new pricing is, to me as someone just enjoying making crazy imagines…

My heart sank when I saw the pricing model. I’ve been creating generative art since 2016 and I’ve been anxiously waiting for my invite. I wont be able to afford to generate the volume of images it takes to get good ones at this price point. I can afford $20/mo for something like this but I just can’t swing $200 to $300 it realistically takes to get interesting art out of these CLIP-centric models. Heck, the initial 5…

MidJourney is a good alternative. Maybe not quite as good as DALL-E, but close enough, without a waitlist and with hobby-friendly prices ($10/month for 200 images/month, or $30 for unlimited)

Re: DALL·E now available in beta

#346

Post by arecurrence »

speedgoose wrote: post #334
smusamashah wrote: post #286

Earlier quoted context omitted.

They won't. DALL-E images are mostly not as high quality. The high quality stuff which everyone has been sharing is result of lots of cherry picking.

In my experience it doesn’t require that much cherry picking if you use a carefully crafted prompt. For example: “ A professional photography of a software developer talking to a plastic duck on his desk, bright smooth lighting, f2.2, bokeh, Leica, corporate stock picture, highly detailed” And this is the first picture I got: https://labs.openai.com/s/lSWOnxbHBYQAtli9CYlZGqcZ It got it a bit strong on the depth of fi…

Additionally, wherever it classically falls over (such as currently for realistic human faces), there will be second pass models that both detect and replace all the faces with realistic ones. People are already using models that alter eyes to be life-like with excellent results (many of the dalle-2 ones appear somewhat dead atm).

Re: DALL·E now available in beta

#347

Post by visarga »

StrictDabbler wrote: post #125
cypress66 wrote: post #35

> Reducing bias: We implemented a new technique so that DALL·E generates images of people that more accurately reflect the diversity of the world’s population. This technique is applied at the system level when DALL·E is given a prompt about an individual that does not specify race or gender, like “CEO.” Will it do it "more accurately" as they claim? As in, if 90% of CEOs are male, then the odds of a CEO being male i…

If accurately reflects the world population then only one in six pictures will be a white person. Half the pictures will be Asian, another sixth will be Indian. Slightly more than half of the pictures will be women. That accurately represents the world's diversity. It won't accurately reflect the world's power balance but that doesn't seem to be their goal. If you want to say "white male CEO" because you want results…

You know a surprising way to solve the issues you presented? You train another model to trick DALL-E to generate undesirable images. It will use all its generative skills to probe for prompts. Then you can use those prompts to fine-tune the original model. So you use generative models as a devil's advocate.

- Red Teaming Language Models with Language Models

https://arxiv.org/abs/2202.03286

Re: DALL·E now available in beta

#348

Post by dplavery92 »

TigeriusKirk wrote: post #252
dave_sullivan wrote: post #248

Earlier quoted context omitted.

I think people don't realize how huge these models really are. When they're free, it's pretty cool. But charge an amount where there's actual profit in the product? Suddenly seems very expensive and not economically viable for a lot of use cases. We are still in the "you need a supercomputer" phase of these models for now. Something like DALLE mini is much more accessible but the results aren't good enough. Early ear…

What are the resources at work here? What are the resources needed to train this model? If someone just gave you the model for free, what resources would you need to use it to generate new results?

In the unCLIP/DALL-E 2 paper[0], they train the encoder/decoder with 650M/250M images respectively. The decoder alone has 3.5B parameters, and the combined priors with the encoder/decoder are the in the neighborhood of ~6B parameters. This is large, but small compared to the name-brand "large language models" (GPT3 et. al.)

This means the parameters of the trained model fit in something like 7GB (decoder only, half-precision floats) to 24GB (full model, full-precision). To actually run the model, you will need to store those parameters, as well as the activations for each parameter on each image you are running, in (video) memory. To run the full model on device at inference time (rather than r/w to host between each stage of the model) you would probably want an enterprise cloud/data-center GPU like an NVIDIA A100, especially if running batches of more than one image.

The training set size is ~97TB of imagery. I don't think they've shared exactly how long the model trained for, but the original CLIP dataset announcement used some benchmark GPU training tasks that were 16 GPU-days each. If I were to WAG the training time for their commercial DALL-E 2 model, it'd probably be a couple of weeks of training distributed across a couple hundred GPUs. For better insight into what it takes to train (the different stages/components of) a comparable model, you can look through an open-source effort to replicate DALL-E 2.[2]

[0] https://cdn.openai.com/papers/dall-e-2.pdf [1] https://openai.com/blog/clip/ [2] https://github.com/lucidrains/dalle2-pytorch

Re: DALL·E now available in beta

#349

Post by cornel_io »

speedgoose wrote: post #334
smusamashah wrote: post #286

Earlier quoted context omitted.

They won't. DALL-E images are mostly not as high quality. The high quality stuff which everyone has been sharing is result of lots of cherry picking.

In my experience it doesn’t require that much cherry picking if you use a carefully crafted prompt. For example: “ A professional photography of a software developer talking to a plastic duck on his desk, bright smooth lighting, f2.2, bokeh, Leica, corporate stock picture, highly detailed” And this is the first picture I got: https://labs.openai.com/s/lSWOnxbHBYQAtli9CYlZGqcZ It got it a bit strong on the depth of fi…

NB: when you share links like that, nobody who doesn't have access can see the results
Post reply on HN