Show HN: Makeai.art – Create Art from a Description(makeai.art)

19 ポイント投稿者 axg115 年前28 コメント

Show HN: Makeai.art – Create Art from a Description

https://makeai.art

28 コメント

For HN: There is and active community on Reddit exploring this tech using free Google Colab notebooks. A few of us pay Google $10/month for practically unlimited usage. https://www.reddit.com/user/corysama/m/gan_art/

For OP: The notebooks have very awkward UX. So, putting a pretty front-end on it is a value-add. But, it won't last. Your base cost is 5 minutes of GPU time. You will very soon be competing with free downloadable programs and much cheaper alternatives. And, I know from a ton of experiments that getting good results out of these systems takes a lot of practice and attempts. Charging $1 per dice roll where most of the time you roll crap will make people angry fast.

What I'd recommend is building a machine with as many 32GB K80 GPUs as possible. They are awkward to build around and quite slow, but cheap if you can find them and have a surprisingly large amount of RAM. Let people use that machine for free to generate low-rez images. Limit 1 GPU per person at a time. And, figure out how to use a low-rez generated image to seed a 1080p version in the GAN using a CPU run with 64GB of RAM. Then take the output of that and AI-upscale it to 4K or something. Charge for the upscaled, denoised version of their favorite creations. Not for every time they roll the dice.

Then hook up with a painting-on-demand service and sell drop-shipped prints of people creations.

axg115 年前

This is an interesting idea, essentially a platform to allow people to produce interesting/creative images and distribute from there.

I agree with you that this is not a giant leap or value-add from what is out there. I'm hoping to iterate from where it is right now. I created this over the last few days as a hobby project to learn a few technologies.

spython5 年前

The K80 seems indeed quite affordable. But would multi-GPU setup work well for these cases? Is the RAM shared?

corysama5 年前

Not shared. But, you want to run lots of times. So, 8 cheap GPUs can run 8 separate experiments in parallel.

GusRuss895 年前

I've been running an AI Art Generator for a few years called NightCafe Creator and we do most of what you suggested here. We run on a managed Kubernetes provider called Coreweave Cloud, that provides nodes with good GPUs for a good price.

From experience I can tell you that fewer people will buy prints than you'd expect (that was my original business idea) BUT more people will pay just to create the art than you'd expect.

I just launched VQGAN+CLIP on NightCafe Creator on Show HN (https://news.ycombinator.com/item?id=28081130) and a commenter directed me here. I'm impressed that the OP here has managed to get this done in such a short time (since VQGAN+CLIP is still quite new).

Since you left such thoughtful feedback here, I'd be keen to get your thoughts on my own version.

hwayne5 年前

It's possible to make the notebook UX a little better using IPython widgets and the form mode. That's what I did for `https://is.gd/artmachine`. You have to strip away a lot of the power to make the UX nicer, but it's worth it for the accessibility.

ronenlh5 年前

I would make all creations public by default, and have a gallery area. This would add to my trust in what the 0.99$ does, as well as be cool on its own.

axg115 年前

For those curious to try but don't want to pay, use the code: MACHINELEARNINGART to try for free (valid for 5 uses).

spython5 年前

I tried to contact you by mail but it got returned with: Relay access denied (in reply to RCPT TO command).

axg115 年前

Apologies - this should be fixed now. If you have any issues email: [email protected]

entropie5 年前

This code is invalid.

axg115 年前

The free codes have been used up unfortunately. I'll add more when the current queue length becomes shorter.

anotheryou5 年前

Are you one of the girls or guys behind this?:

https://colab.research.google.com/drive/1Foi0mCSE6NrW9oI3Fhn...

It's basically that in monetized/with stripe, right?

axg115 年前

It's a very high-level architecture to that, yes :) I did some fine-tuning on scraped images of modern art to hopefully result in more visually pleasing images.

Malcx5 年前

$1 per image? That's going to be a really tough sell based on the quality of the sample images. Without a free trial I'll never get to see if this is any good for my own stuff.

("free" means no account creation needed even if the output is watermarked)

axg115 年前

That's a great point regarding a free watermarked trial. I'll look into adding that.

armchairhacker5 年前

I would pay for a service like this, but $1 per image is too much. Does it really cost that much to run the neural network and create 1 image?

axg115 年前

It's surprisingly expensive to compute, not quite $1 but after stripe fees it's not far off. I'm looking at ways to bring the compute cost down.

GusRuss895 年前

I use Coreweave Cloud for NightCafe Creator (essentially the same service as you, but more mature). Check it out.

balazshoranyi5 年前

A pretrained model from the good folks over at huggingface: https://huggingface.co/spaces/flax-community/dalle-mini. Also provides a free demo.

axg115 年前

Huggingface is really an incredible org. I hope they continue on the same excellent path. They recently raised $40M and I imagine there will eventually be stronger pressure to monetize their work.

protogenes5 年前

I did something similar - free to use for image generation but you can get your own artwork on Canvas:

https://neuralblender.com/

axg115 年前

Great, I like the name! I'm trying it now.

axg115 年前

Here's my result:

https://s3.amazonaws.com/ai.protogenes/art/3ed8a7fa-f60b-11e...

"a monkey riding a rocket"

I think it works well :) Which architecture is this based on? Did you consider any trips or fine-tuning to generate images that are perceptually similar to art?

protogenes5 年前

It is based on VQGAN (trained on imagenet) and Open AI's CLIP. To make it more similar to art you would need to do some "prompt engineering. Try: "painting of ..." or "... painted by Van Gogh"

protogenes5 年前

to everyone who is trying it -> just scaling up the system now. Please be patient and keep your tab open. The image will eventually appear :)

corysama5 年前

Clicking "Get it on canvas" takes me straight to "enter your personal information" with no clue of what's going on. Noooooooope.

protogenes5 年前

Thanks for pointing this out. Actually NeuralBlender.com was not to be released to the public yet, but the topic in this thread kind of invited it. I will clarify that and improve the UX.