60 karma · joined March 10, 2015
NightCafe has a feed of artworks, and I too was getting a lot of NSFW or very distasteful prompts. I just used an npm package to detect 'naughty words' to auto-flag any distasteful ones as 'needs moderating', where I can then manually approve or reject them.
However, if you're wanting to sell digital products, it may not be the best option for you. I've had to do a bit of extra work to get it to work for my situation.
Style transfer on NightCafe has a bulk-create mode that has been used in the past to create 500+ creations at once and they all finished in 10 mins.
Most of my regular users are artists who either use it for inspiration or use it to mix into their digital art. (NightCafe has been around for a while as a style transfer app).
It's rare that people buy prints. Most money comes from people buying credits.
I won't claim to have a good knowledge of the inner workings, but basically CLIP will give a score for how well the image matches your prompt, and that's used as a loss function for the GAN.
On the Colab notebooks you can specify multiple prompts with different weights, in which case I assume it has multiple loss functions that it's trying to optimise for.
I'm more of a web app guy than a deep tech guy. Some of what I wrote can probably be corrected by someone with a better knowledge of the ML.
I will build an interface for it rather than a prompt syntax. My target audience is a bit less technical than the average Google Colab user.
However, adding keywords like "watercolour painting", "van gogh painting" etc goes a long way to getting more precise results.
One big difference between style transfer and GAN art is that style transfer never changes the shape of things. E.g. you can't put in a woman's face and a cubist painting and get out a rearranged version of the face (as cubists do). It will have all the hallmarks of cubism, but it's still recognisable as the original photo.
With GAN art though, you can put that photo in as a start image, say "Cubist painting of a woman's face", and the GAN will actually rearrange the face like a cubist would.
I almost gave up on NightCafe when I was trying to make money from selling prints to people. Then, one user generated 3,000 creations in a week, which cost him nothing and me $300. I decided to spend the weekend implementing a credit system to prevent abuse, and was surprised to find that more people were willing to pay for credits than I expected.
It's still not much and I've had to spend a lot of effort on reducing costs in various ways just to be able to break even. But I love working on this app and have high hopes that _one day_ it will become profitable enough for me to work on it full time.
My best tips for good results - use a phrase describing objects you'd commonly see in pictures and art, like "House on a hill" or "Futuristic city", add some modifiers, then start with the smallest resolution and runtime. Try a few variations of phrase and modifiers, and see what looks promising. Then pick the best one(s) and use the "Evolve" button to run them for longer and increase the resolution.
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.
Looks like they're both new apps, so they've done well to get everything up and running so quickly! I already had a lot of the groundwork laid for NightCafe Creator, so it was much easier for me to get this done. There's a lot involved, especially if you're going to charge for use (which is a must if you expect to have any sort of volume in the long run).
Advantages over using Google Colab:
- It's >2x faster (runs on NVIDIA RTX A5000 GPUs). E.g. 800 iterations at 600x600px in 8 mins vs 19 mins on Colab.
- Can do higher resolution (894x894px > 700x700px).
- Run multiple jobs in parallel.
- Works from any device.
- Obviously it's much easier to use for anyone non-technical.
- It's easy to "evolve" creations. I.e. tweak the prompt, increase the resolution, and/or just run them for longer.
- You don't need to keep the tab open to see the result.
It's free for basic use, but does cost money for extended use. Unfortunately I don't have the same access to GPUs as Google does.
Try it out. Would love to get the thoughts of the HN community!
It's far easier to integrate than the current go-to virtualization options, which often require you to re-architect a lot of how your list works.
Thanks for trying it out!
I've thought a bit about marketing to Instagrammers, I think neural style for short videos would be a good service to offer. Not sure where it fits on my roadmap yet though.