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

Stable Diffusion 2.0

stability.ai

Post reply on HN
Print view View original

141–150 of 519 posts

Re: Stable Diffusion 2.0

#141

Post by Jack000 »

hackater wrote: post #130
Jack000 wrote: post #84

Earlier quoted context omitted.

Depends on the dataset. You can probably get decent results by restricting the modality of the images (faces, cars, bedrooms etc) I trained from scratch with 4x3090 and while it’s not as good as SD it’s surprisingly better with hands.

Can you go into a bit more detail? What architecture did you use? Is the month training time really just training with mini batches with a constant learning rate? Or are these many failed attempts until you trained a successful model for a few days in the end? I particularly interested in the image generation part (the DDPM/SGM)

Yeah I did have a few false starts. Total time is more like 3 months vs 1 month for the final model. For small scale training I found it’s necessary to use a long lr warmup period, followed by constant lr.

There’s code on my GitHub (glid3)

edit: The architecture is identical to SD except I trained on 256px images with cosine noise schedule instead of linear. Using the cosine schedule makes the unet converge faster but can overfit if overtrained.

edit 2: Just tried it again and my model is also pretty bad at hands actually. It does get lucky once in a while though.

Re: Stable Diffusion 2.0

#142

Post by ftufek »

coldblues wrote: post #137

I just thought about this, so bare in mind that I don't know much of the technical implications of this, but: Couldn't we train a very good model by distributing the dataset along with the computing power using something similar to folding@home?

You would likely be limited by the communication latency between nodes, unless you come up with some unique model architecture or training method. Most of these large scale models are trained on GPUs using very high speed interconnects.

Re: Stable Diffusion 2.0

#143

Post by randyrand »

mysterydip wrote: post #21
Gigachad wrote: post #16

Earlier quoted context omitted.

Now that's a can of worms I don't think anyone wants to open.

Some do, that's the problem.

Artist have been drawing people of all ages having sex for literally thousands of years. Why should I care about that?

Re: Stable Diffusion 2.0

#144

Post by ftufek »

tamersalama wrote: post #107
LASR wrote: post #27

I am a solo dev working on a creative content creation app to leverage the latest developments in AI. Demoing even the v1 of stable diffusion to the non-technical general users blows them away completely. Now that v2 is here, it’s clear we’re not able to keep pace in developing products to take advantage of it. The general public still is blown away by autosuggest in mobile OS keyboards. Very few really know how far…

I agree that this a big wave, but I'm still struggling to find commercial (read: large organizations) applications.

I suspect those applications will come from specializing the model. For example, there's people that have avatar generators or automated ad creatives. A cool application I've been toying with is generating icons.

Re: Stable Diffusion 2.0

#145

Post by __rito__ »

kmeisthax wrote: post #58

Is there a good explanation of how to train this from scratch with a custom dataset[0]? I've been looking around the documentation on Huggingface, but all I could find was either how to train unconditional U-Nets[1], or how to use the pretrained Stable Diffusion model to process image prompts (which I already know how to do). Writing a training loop for CLIP manually wound up with me banging against all sorts of stra…

Well, you can learn about generative models from MOOCs like the ones taught at UMich, Universitat Tubingen, or New York University (taught by Yann LeCun), and can gain knowledge there.

You can also watch the fast.ai MOOC titled Deep Learning from Scratch to Stable Diffusion [0].

You can also look at open source implementation of text2image models like Dall-E Mini or the works of lucid rain.

I worked on the Dall-E Mini project, and the technical knowhow that you need isn’t closely taught at MOOCs. You need to know, on top of Deep Learning theory, many tricks, gotchas, workarounds, etc.

You could follow the works of Eluther AI, follow Boris Dayma (project leader of Dall-E Mini) and Horace Ho on twitter. And any such people who have significant experience in practical AI and regularly share their tricks. The PyTorch forums is also a good place.

Learn PyTorch and/or JAX/Flax really well.

[0]: https://www.fast.ai/posts/part2-2022.html

Re: Stable Diffusion 2.0

#146

Post by wodenokoto »

tamersalama wrote: post #107
LASR wrote: post #27

I am a solo dev working on a creative content creation app to leverage the latest developments in AI. Demoing even the v1 of stable diffusion to the non-technical general users blows them away completely. Now that v2 is here, it’s clear we’re not able to keep pace in developing products to take advantage of it. The general public still is blown away by autosuggest in mobile OS keyboards. Very few really know how far…

I agree that this a big wave, but I'm still struggling to find commercial (read: large organizations) applications.

Built a plugin for Power Point and sell it corporate wide.

Re: Stable Diffusion 2.0

#147

Post by eugenhotaj »

coldblues wrote: post #137

I just thought about this, so bare in mind that I don't know much of the technical implications of this, but: Couldn't we train a very good model by distributing the dataset along with the computing power using something similar to folding@home?

The network communication overhead would be way too high to make this useful. At least for current methods of training large models.

Re: Stable Diffusion 2.0

#149

Post by prawn »

fassssst wrote: post #136
prawn wrote: post #131

Hopefully related: If I'm a photographer wanting to improve resolution of my content for printing, what's my current best bet for upscaling? Is it realistic to make use of this on the command line, feeding it my own images? Or has someone wrapped it in an app or online service?

Photoshop and Lightroom have had AI upscaling for awhile.

Ah - had forgotten. I'll try them first. Thanks.

Re: Stable Diffusion 2.0

#150

Post by version_five »

Gigachad wrote: post #11
argella wrote: post #6

Earlier quoted context omitted.

Bummer. AI porn is fun.

The future is probably models trained almost exclusively on porn.

No. The whole point of these models is that they combine information across domains to be able to create new images. If you trained something just on, say baseball, you could only generate the normal things that happen in baseball. If you wanted to generate a picture of a bear surfing around the bases after hitting a home run, you'd need a model that also had bears and surfing in the training data, and enough other stuff to understand the relationships involved in positioning everything and changing poses.
Post reply on HN