Skip Navigation
Remote View
test @sh.itjust.works
tal @lemmy.today

Test post to try out db0's aihorde bot

db0 apparently has a bot, @aihorde@lemmy.db0.com, linked up to AI Horde, which will render images using various LLM models in response to comments.

All using it requires is, in a comment, writing the text: 

 undefined
        @aihorde@lemmy.db0.com draw for me <prompt text>

It's apparently been active for years, but somehow, I'd been unaware of it until now, when db0 happened to trigger it in a thread on !imageai@sh.itjust.work where we were creating images to represent usernames. It looks like nobody's been aware of it, and it's gotten very little use.

It apparently supports multiple models (including Pony, which is trained on and can render furry art, and Flux, which does pretty well with ordinary, English-language descriptions). It looks like the default is Flux.

Note that according to the FAQ, it will try to block (and will do a 1-day tempban on users for) NSFW content.

I'm going to try out a few prompts here and maybe mention this somewhere higher-profile, because this is really cool --- it lets users pseudononymously generate images on the Threadiverse without needing to use a commercial service or having their own hardware.

Bot FAQ:

https://wiki.dbzer0.com/cbotart/

AI Horde style list (these can be accessed by adding style: <stylename> to the end of a prompt). Some of these are clearly different models whose names I recognize, and I'm pretty sure that some are just adding some extra text to the prompt.

https://github.com/amiantos/AI-Horde-Styles-Previews/blob/main/previews.md

While Lemmy doesn't natively provide the ability to follow a user, it posts what it generates to !botart@lemmy.dbzer0.com, so you can subscribe and watch generation if you want.

63 comments

  • Feel free to play as much as your want, everyone . If you like this service, consider onboarding your own gpu to the horde 😉

  • To repeat the first test post I ran (just to see whether it gives the same output, which would mean that it uses a fixed seed, or whether it gives different output, which means that it uses some non-deterministic process):

    @aihorde@lemmy.dbzer0.com draw for me An engraving of a skunk. style: flux

    • Here are some images matching your request

      Prompt: An engraving of a skunk.

      Style: flux

      • Okay, it differs, so that probably means that it's set up to use a random seed. Hmm. That's not my preference, since it makes it hard to narrow down and refine a precise image, but some people seem to like that as a way to generate more images.

  • Now I just want to confirm that the default style is Flux; I think it is, from db0's comments. That would be good, because Flux is pretty easy to use and can just take what more-or-less amounts to English-language prompts. Many other models want something more like a list of keywords.

    @aihorde@lemmy.dbzer0.com draw for me A delicious slice of cheesecake with a delicious scoop of strawberry ice cream on the side and chocolate shavings sprinkled on top. The cheesecake is on a white plate.

  • Some time back, I rendered some landscape paintings in the style of Casilear. I don't know whether the bot can do different-resolution images --- I did those at 2560x1440, whereas the bot is using 1024x1024 in what I've seen, and I don't see a reference in a quick search of the docs to a way to change that. I was also using a model, realmixXL_v15, derived from Stable Diffusion XL but not itself Stable Diffusion, and it doesn't look like AI Horde has that model available. So I don't know how well this will work. But let's see what happens if we ask it to regenerate such an image using SDXL:

    @aihorde@lemmy.dbzer0.com draw for me Scotts Bluff, by John William Casilear style: sdxl

    • Here are some images matching your request

      Prompt: Scotts Bluff, by John William Casilear

      Style: sdxl

      • Great! Okay, those have very low saturation, look almost monochrome. The bot apparently supports use of a negative prompt --- this is where one specifies keywords that one wants excluded rather included in an image --- if you append them after three pound signs in a row. Flux doesn't support negative prompts, but Stable Diffusion does. Let's see if we can get some color by making use of that functionality:

        @aihorde@lemmy.dbzer0.com draw for me Scotts Bluff, by John William Casilear###monochrome, desaturated style: sdxl

  • Now, let's try out another model, just to make sure that styles that entail having a model downloaded to a node actually work with this. Looking at the bot's history, it looks like it has been used to render things in other styles, but I don't know for sure whether those are styles that involve use of another model.

    Pony Diffusion doesn't really take English-language prompts (though it'll probably do something not-entirely-unreasonable with them); it really expects a list of keywords. However, it does understand Danbooru tag groups:

    https://danbooru.donmai.us/wiki_pages/tag_groups

    @aihorde@lemmy.dbzer0.com draw for me score_9, score_8_up, score_7_up, cat_girl, business_suit, morning, drinking coffee style: pony

    • Here are some images matching your request

      Prompt: score_9, score_8_up, score_7_up, cat_girl, business_suit, morning, drinking coffee

      Style: pony

      • Man, that ran quickly enough that whatever rendered the thing had to have had Pony Diffusion either already downloaded or I'm misunderstanding how AI Horde functions...maybe it normally farms out images to the cloud and is able to intelligently try to find a node that already has a given model downloaded rather than reloading a new model or something.

        Okay, a repeat, using Nova Furry Pony:

        @aihorde@lemmy.dbzer0.com draw for me score_9, score_8_up, score_7_up, cat_girl, business_suit, morning, drinking coffee style: nova furry pony

  • Let's see how well some past combinations of prompt terms that I spent a while banging together on Stable Diffusion XL or derived models to get the effect of specific styles work.

    @aihorde@lemmy.dbzer0.com draw for me a man standing on a rainy street, noir, vector illustration, ink, hard light, light on dark,black-and-white,shadowed style: sdxl

    • Here are some images matching your request

      Prompt: a man standing on a rainy street, noir, vector illustration, ink, hard light, light on dark,black-and-white,shadowed

      Style: sdxl

      • Fantastic, so that works. I'm a sucker for high-contrast black-and-white stuff. Another set of noir prompt terms that I've liked; again, dunno if the base SDXL model has sufficient training to do it:

        @aihorde@lemmy.dbzer0.com draw for me a man standing on a rainy street, noir, by darwyn cooke style: sdxl

63 comments