Yes I know this was asked before: https://lemmy.world/post/8258092

And it seems back then the model used might have been SD1.5+Deliberate V2

Possibly this one : https://huggingface.co/XpucT/Deliberate/tree/main

But results don’t match.

EDIT : By this I mean if run an image prompt for a given seed on any Deliberate model with settings that match the percance generator then I get very different results

It would be nice to have the perchance SD 1.5 model uploaded on huggingface so people can use it on other platforms and/or privately.

The perchance text-to-image generator is a really well-balanced SD 1.5 model!

I feel it would be good for the SD community to have it available for download online.

  • @AdComfortable1514OPEnglish
    1•1 month ago

    Added an EDIT , to give context to “results don’t match”.

    //–//

    Yeah , that sounds reasonable. Thanks for pinging the dev.

    But however I look at it , the Perchance model can’t be the Deliberate V2 model.

    I just want the .safetensors file of the perchance SD1.5 model to be available online.

    • @perchanceMEnglish
      3•1 month ago

      As mentioned a few times elsewhere on lemmy, text to image plugin currently uses a few different top models on civit depending on they keywords in the prompt. I don’t make a habit of mentioning specific models used for AI plugins mostly because some of the top-performing OSS models (even on SFW content) have NSFW names/reputations (I use some regex adjustments to “tame” these sorts of models). There are actually 2 other reasons that it’s not a great idea for me to mention specific models names, which I’ll leave as an exercise to the reader[1]. There’s really no benefit to me naming the models because anyone who really wants to know the specific models used will have enough motivation to simply test the top models on civit (or localllama for text gen). It’s not hard. When I say “top models on civit” it narrows it down to like 30 mins of work - if they’re a fan of the Perchance platform, then they’ll understand that this trade-off (30 mins of work for them, for lack-of-several-headaches for me) is a good one.

      [1] And to be very clear: This isn’t about a “competitive advantage” or anything like that. Anyone “competing” with Perchance (in whatever sense), but unable to do 30 mins of work to test a handful of top OSS models is… gonna have a hard time.

      • @AdComfortable1514OPEnglish
        2•1 month ago

        Yeah , I understand the situation.

        I appreciate you telling me model selection is within the top categories on civitai.

        I didn’t know multiple models were being used.

        Any chance you could allow a user to specify which model to generate with , to override switching models based on keywords?

        Like for example writing (model:::Anime) in the prompt to use an anime model?

        • @perchanceMEnglish
          2•1 month ago

          I think the multi-model situation is temporary, and won’t be needed for the next generation. I could still do something like that and keep it backwards-compatible-ish, but next generation is coming quite soon so probably not worth it (we’ll see though)

          • @AdComfortable1514OPEnglish
            1•1 month ago

            I’d say yes

            Add a little code snippet that explicitly selects the model.

            If a new update is coming “quite soon”

            , then that could be several months from now.

            And if the new update means

            by default replacing the current (three?) SD 1.5 models

            with a single SD 1.5 model.

            Why not use all of them? As in older models + the new one , via this system?

            //—//

            So what I mean by this is;

            In the next update ,

            The new SD 1.5 model could be available by default

            , and if user writes (model:::anime) , (model:::photo) or (model:::anthro)

            , then it uses the “older” SD 1.5 models.

        • VioneTMEnglish
          2•1 month ago

          Second this, so we can probably use ‘anime’ keywords in a ‘realistic’ model :)