Several months ago I said new image generation quality would be coming soon. Then flux came out and I was like “oh cool, i’ll just wait a month or so for community finetunes”, and once again informed people that an image gen upgrade was not far off. But it turns out flux is really hard to finetune in its current form.

Aside: There have been attempts to fix this issue, but we’re not quite there. I’ve been helping some people who are working on this (mainly dataset stuff, I’m no ML researcher), and progress is being made, but we’re still at least a month away from ‘serious’ flux finetunes.

So base flux still doesn’t know ‘basic’ stuff (e.g. doesn’t even know most pokemon), and illustrious (another new model) requires a very specific prompting format.

While helping with the eventual open source flux finetune, I have also been attempting to put together a system that would intelligently route to the best model based on prompt content, and also generate tags for illustrious based on a natural language prompt, but it’s still not good enough.

So Perchance images are atrociously bad at this point, and I considered just upgrading to SDXL, but this would likely mean two upgrades in a short period of time, both of which would require prompt engineering adjustments on the part of perchance generator devs. That would be annoying, and maybe more painful than just dealing with bad generation quality for another month or two.

In hindsight, I should have just upgraded to SDXL midway through 2024 (or even earlier). We may actually get another text gen upgrade before the image gen one at this rate. We’re also getting close on video gen now with models like HunyuanVideo, which seems to be finetunable, and is quite fast with FastVideo.

Tangentially, I’ve been spending a lot of time on behind the scenes server stuff recently. For example, I’ve had to add filters to prevent people from uploading literal CSAM to perchance.org/upload - a problem that I naively did not consider when first creating the upload feature. This sort of work is annoying because it doesn’t result in fun new features or plugins, but spending time on automating this sort of thing is important, because it ensures that e.g. using features like /upload doesn’t require logging in, and doesn’t e.g. require employing people for moderation. I’d much rather move a bit slower, and ensure perchance’s sustainability and complete independence.

And tangential to that: One thing that I want to publicly promise, just so I can say “I told you so” in 20 years, is that Perchance will never “sell out” or “rug pull” in any sense of either phrase. It’ll always be a bit weird. It’ll never get investors, I’ll never sell it, never require login, never send you emails (except e.g. password reset), never put ads on generators (unless it imports an AI/server-GPU-powered plugin), never add user-hostile social mechanics that try to increase ‘engagement metrics’, and so on. The OG devs here know this I hope, but there are newbies and non-devs here who think perchance is just another “AI site” that is burning investor money to keep it free, in preparation for a rugpull once they have market share. Perchance is a different kind of website. It’s a public good that I maintain, not a “startup”. The price you pay for this as someone who uses perchance is slower development, which I think is worth it, especially considering that it’s always been like this, and people seem to like perchance (though I’m sure many wish I could fix/improve things faster).

So anyway, this was (supposed to be) just a quick post about what’s been happening recently on the dev side of things. Apologies for the huge delay on the image gen side of things. Also sorry for the lack of response to a lot of posts and messages - I have a large backlog of stuff to get to (as usual, please feel free to ping me again and/or repost weekly).