Hi there, while the word list will help, I think plugging an LLM to a user-facing form is still a very, very bad idea, especially if this is meant for children.
I tried "stegosaurus with giant lollipop", but I got a tyrannosaurus with three stegosaurus legs and stegosaurus spines but no thagomizer and no lollipop: https://imgur.com/a/1EPUm3F. When I tried again it seemed to have frozen. The third try gave me the second tyrannosaurus mutant hybrid, but I did get a giant lollipop.
I agree, no-signup is really nice, esp. since I'm hesitant to have my kids signup for things, and of course free is nice. That said, if you do want or need to monetize, one idea would be to slap a donation link like 'Buy me a beer' on it. I could see some people being willing to send a few bucks to your daughters' college fund in exchange for infinite coloring pages for their kids. Also, it probably makes sense to cache if you're not already - once someone generates "Paw Patrol vs Scooby Doo", put it in CDN and serve up from there instead of generating again.
If the costs were to increase enormously, I'm thinking of limiting the generation via machine id, or that the person would have to buy credits or something like that.
Thanks for the CDN tip, will look into it. Currently hosting via Cloudflare, so shouldn't be that hard I guess.. (Famous last words)
Doesn't make it legal.
As a small site, I'm sure you won't draw the ire of some big companies, but if you start charging money, you'll likely get a cess and desist.
This is a fun idea, but whatever version of stable diffusion being used isn’t very prompt adherent so any kind of complex scene or interaction between characters is mostly ignored, which reduces the fun a bit.
Whenever I show things like this to my kids they always say something tbat’ll be really hard - like ‘a unicorn riding a princess’ and then everything comes back as princess riding a unicorn and they say ‘this sucks’
Trope subversion is always difficult since there is naturally so little training data in most checkpoints (SD, XL, etc.) that reverses things like this (mermaid with human legs and fish head, Cerberus with five heads, a piano where natural keys are black and the sharps/flats are white, etc.)
Outside of manual control like ControlNets / Inpainting or a custom LoRa, there's not much you can do except "re-roll" hoping you'll get lucky.
I wonder if when language models are better integrated with image generation models it’ll get any better at this. Or is this a fundamental issue that can’t be solved - it’s not like we’re going to add these edge cases to the training data
This is fantastic. Can you share any details of how you created the pictures?
I am trying to do illustrations for a childrens book in a similar style (although I want much more detail on the pages) but all my attempts end up being a complete mess.
But even if you can't share deets thanks for the inspo!
It is most likely using a Lora - it’s like an add on for stable diffusion that forces a specific style. You can find them ready-made for all sorts of styles - including black and white line art. Or you can also train your own using a few examples of a style you want to use.
Love the concept. Though every image is pretty dystopian
Evil monkey claw https://www.coloringsai.com/en/coloring-page/a-christmas-the...
Floating spiderman head: https://www.coloringsai.com/en/coloring-page/spider-man-and-...
Whatever this is: https://www.coloringsai.com/en/coloring-page/the-dark-depths...
Raccoon head on baby's body: https://www.coloringsai.com/en/coloring-page/a-diapered-racc...
Bear cub with wheels: https://www.coloringsai.com/en/coloring-page/a-bear-family-c...
Raccoon head on baby’s body is very cute
Seems like HN is having fun with it, not sure I want a 4-year-old coloring any of these:
- https://www.coloringsai.com/en/coloring-page/multiple-africa...
- https://www.coloringsai.com/en/coloring-page/a-penis-cm406l5...
- https://www.coloringsai.com/en/coloring-page/beutiful-girl-i...
- https://www.coloringsai.com/en/coloring-page/bloody-inhumane...
- https://www.coloringsai.com/en/coloring-page/the-body-of-a-b...
- https://www.coloringsai.com/en/coloring-page/the-torso-of-a-...
Seems like a normal TTP.
https://knowyourmeme.com/memes/time-to-penis-ttp
Thanks for checking out the app. You're absolutely right, although that's still hard to stop with text input I guess..
I already use OpenAI's moderator tool to block the worst and am working on a blacklist of words to block.
Luckily I still choose the coloring pages for my daughter myself ;
The AI title and descriptions are lovely too:
> https://www.coloringsai.com/en/coloring-page/10-10-beautiful...
> Sexy Cartoon Woman Coloring Page
> A coloring page of a woman in a cartoon style, with adult themes, ideal for kids who love drawing sexy characters.
Hi there, while the word list will help, I think plugging an LLM to a user-facing form is still a very, very bad idea, especially if this is meant for children.
I tried "stegosaurus with giant lollipop", but I got a tyrannosaurus with three stegosaurus legs and stegosaurus spines but no thagomizer and no lollipop: https://imgur.com/a/1EPUm3F. When I tried again it seemed to have frozen. The third try gave me the second tyrannosaurus mutant hybrid, but I did get a giant lollipop.
Thanks for checking out the site!
I'm currently using a cheap model, so there's a good chance the rate limiter will be hit with the traffic from HN going to the site.
I didn't expect it to be a success, but it's being worked on :)
There's a couple of these out there already - but the lack of a sign-up requirement is a nice touch.
https://www.coloringbook.ai
https://colorbliss.com
I agree, no-signup is really nice, esp. since I'm hesitant to have my kids signup for things, and of course free is nice. That said, if you do want or need to monetize, one idea would be to slap a donation link like 'Buy me a beer' on it. I could see some people being willing to send a few bucks to your daughters' college fund in exchange for infinite coloring pages for their kids. Also, it probably makes sense to cache if you're not already - once someone generates "Paw Patrol vs Scooby Doo", put it in CDN and serve up from there instead of generating again.
Thanks for thinking along, that's a good idea!
If the costs were to increase enormously, I'm thinking of limiting the generation via machine id, or that the person would have to buy credits or something like that.
Thanks for the CDN tip, will look into it. Currently hosting via Cloudflare, so shouldn't be that hard I guess.. (Famous last words)
edit: Created the buyMeACoffee :) https://buymeacoffee.com/coloringsai
Maybe "Buy us some crayons" instead? You could even do a partial donation for coloring supplies to a kids' charity.
Thanks!
Neat. What model are you using? https://huggingface.co/artificialguybr/ColoringBookRedmond-V... or one of these https://civitai.com/tag/coloring%20book ?
Thanks for sharing, I am not using any of those. Willing to train my own model in future as well!
Any issue with copyright? Lots of brands named right in your copy.
Yea, OP couldn't have picked worse companies for aggressive copyright protection.
I used to work at Disney, and was amazed that they’re so large that their companies sue each other. Disney is so litigious that it sues itself lol.
Thanks for your thoughts. To be honest, I have no idea.
But if I google "Coloring Pages Pokemon", for example, I find dozens/hundreds of websites that are not Pokemon, so to speak.
Doesn't make it legal. As a small site, I'm sure you won't draw the ire of some big companies, but if you start charging money, you'll likely get a cess and desist.
This is a fun idea, but whatever version of stable diffusion being used isn’t very prompt adherent so any kind of complex scene or interaction between characters is mostly ignored, which reduces the fun a bit.
Whenever I show things like this to my kids they always say something tbat’ll be really hard - like ‘a unicorn riding a princess’ and then everything comes back as princess riding a unicorn and they say ‘this sucks’
Trope subversion is always difficult since there is naturally so little training data in most checkpoints (SD, XL, etc.) that reverses things like this (mermaid with human legs and fish head, Cerberus with five heads, a piano where natural keys are black and the sharps/flats are white, etc.)
Outside of manual control like ControlNets / Inpainting or a custom LoRa, there's not much you can do except "re-roll" hoping you'll get lucky.
I wonder if when language models are better integrated with image generation models it’ll get any better at this. Or is this a fundamental issue that can’t be solved - it’s not like we’re going to add these edge cases to the training data
This is fantastic. Can you share any details of how you created the pictures?
I am trying to do illustrations for a childrens book in a similar style (although I want much more detail on the pages) but all my attempts end up being a complete mess.
But even if you can't share deets thanks for the inspo!
It is most likely using a Lora - it’s like an add on for stable diffusion that forces a specific style. You can find them ready-made for all sorts of styles - including black and white line art. Or you can also train your own using a few examples of a style you want to use.
Thanks!
Watch out for Pokemon usage they are super strict about that
Nice project =)
I like to have fun tho...
Yuna baastutions? Original prompt attempt. https://imgur.com/a/WKSWsWT
Not very good at swastika https://imgur.com/a/fD3Y8BZ
But Hitler looks okay https://imgur.com/a/tzLEedR
What a brilliant idea!
The longer an AI app exists, the probability that an output involving Hitler exists grows to one.
- Godwin’s AI Law
I typed "a panopoly of penises playing pirates" and hit "generate" and it seems to be stuck now.