I don't understand why the change in results. used to be it was possible to create wonderful painting styles, realistic, Beaux-Arts, academic, vintage pulp, Art Deco or realistic looking old photos etc and now the results are so childishly cartoonish as to be useless. The same as on any other run-of-the-mill AI generator.

I simply do not understand why the wonderful quality that set Perchance apart is seemingly abandoned. The update that was implemented begin August 2025 was superb. Why degrade now?

Is this a temporary thing or a long-lasting change? Leave well enough alone.

I'm a simple user with very simple prompts (no more than a couple dozen words), I had a great professional photo prompt I used alot but now all the professional (and casual) photos are coming out as cartoons. I'm nowhere near the rest of these people's level of creating, so I don't know what this "limit" is, but I just know I'd like to be able to create realistic images again please.

edit: As an update, now over half my generated images are appearing as if my monitor has been smashed, it does that wierd digital abstract.

I have to say, the only benefit ive experienced from this update was now text renders better. But every other aspect is worse.

The worst seems to be the simple lack of clarifying news about the update. People on reddit are confused and wondering what is going on.

Just let us know if this is permanent and what the changes were so we can adjust, or let us know if we can relax and things will be fixed at some point.

The drastically reduced token count is ruining generators that use style mixing. We used to be make to make amazing unique images, now things seem to default to cartoons.

the style tags on many generators used to be able to mitigate the tendency for the images to skew toward cartoons, but they no longer seem to work.

Weights are totally broken, no more boosting terms to even things out or control the prompt better.

we can still make photo-real images but we can make complex scenes as we have to dedicate so much of the prompt to the quality/camera terms to make the image not be a cartoon. If weights worked we could try to boost a smaller amount of tokens and have more room for the actual prompt we want to achieve. The token count is so small it fights us every step.

So please, just give us some news about what your intentions are!

PLEASE!

+1 I'm getting the results I was getting a year ago with similar prompts. It's not just the issue of photorealistic quality or getting the prompts to work right, I also get quality artefacts I have not been seeing for over 6 months. It seems the previous model was custom-trained, and that this change lost all the ability and quality that have been acquired. Personally, I have seen zero improvement in any area over the last one. And it's not a matter of prompt length.

i used Anime Screencap style but they now less detailed and background almost not exist.

This may explain some of my issues. I was creating images for a gritty 1800s sailing piece, and when I moved to the sequel, suddenly there seemed to be a stifling amount of censorship. In addition, "anti-description" is completely ignored. "Gritty" is completely missing in some cases. Cartoonish. Yeah, that tracks.

For instance, when I say that the pirate sailors are bathing in the altogether, and one of them is a teenager, it's the 1800s, and that's what happens. NOT the sailors are sitting around the deck, fully clothed and with Navy hats.

It seems that the generator is being updated, It is not final yet, and even if it was, we got all that discussion on the previous update that happened last year, people got used to the new images and that was it

I'm also having issues with previous prompts becoming cartoonish/anime out of nowhere, I couldn't find a way to fix that yet, but I dont think there is much we can do except wait.

I wouldn't expect a free tool to have 2 models at the same time

https://lemmy.world/post/41971465

There's currently what I presume is a temporary issue where elements beyond the 250-300th word will be ignored, I'm not 100% certain on what the exact cutoff point is, it seems to vary.

So: I just stripped my prompts a little, mostly removing the redundant and overly verbose descriptions that used to be necessary -- that fixed it right away.

I'm getting downright amazing results with these smaller inputs now, for both realistic and stylized images. The trick really is that you have to put the most important details first, and the least important last. That way you guarantee that the "essence" of the image you want doesn get lost if you inadvertently cross the 250-300 word boundary.

Also, some people are saying that removing negative prompts seemed to help. I can't confirm that because I wasn't using them.

Now, and I'ma go a bit offtopic here, I don't know if this situation is only temporary or what, but it took such a miniscule effort on my part to adapt to the change that, with all due respect, I'm left once again wondering what in Dagon's unholy piehole is wrong with folks rushing to the torch and pitchfork.

Well, business as usual: another update, another riot ;)

Despite the negative feedback I still think this is an epic boost, not a downgrade

Didn't it get posted that the devs are putting more resources towards the Story telling AI and it's making the image generator worse r something? I'm having the same issues with more complex and detailed prompts. Like if I want to make an image that looks like a poster with folds and wrinkles and wear etc, it just wont do it now. but for more simple prompts, like getting an image that just looks like a clean photo. So it's fine for now. Like, I can get the character design, their clothing, and the setting in. But extra stuff, not so much.

Yea its a big downgrade simple images are fine but i can't see any improvements there is nothing this version can do that the old one can't just more limitations. some people mention text on the images is better but this is pointless because 1 i know how to do add that stuff the old school way. and 2 with the hard token Limit if you do that it will probably fuck up the art i am hopeful this will be fixed

It broke a ton of technical and formula related prompting systems as well. Also, fundamentally breaks a lot of novel forms of style blending. On my end it destroyed RltDif as it was being booted up into research, showing massive potential for Reality Environment building. The only other known means of building true Environments is formula prompting. RltDif basically once finished would have offered a Reality Environment fine-tuning dashboard prompt, the new update if this is truly how it is going to be from now on & not just some temporary maintenance awaiting resource allocations adjustments, or rather if it was done becauae of what rudbo said "for better text generation" you are nuts, you can literally select the text with {...} Which keeps it from being affected by noisy generation.... Anways this sets back RltDif and a load of other features that where being worked on. Along with Formula blocks are now functioning kinda wonky, but im sure those will be the first of things I can fix / make what ever changes within formatting for conversions, I don't know how much of everything will still be functional. This update as far as I can tell actually removes a lot more then it adds, that's if it even adds anything at all to begin with LOL. I was about to "add that and 10 to 20 x whatever this terrible over pervasive Style is, & likely a way to have this style and siht "Any style feasible?".. I also had the text generation issue plenty fixed more then fixed I had a whole library of siht for text and creative font generation with out much of an error in text generation.. If this is all for text or something, then why not just limit it and all styles like it to the critique "screencap" or something,.. cause this is overkill ridiculous, not only do some of us not care for perfect paragraphs vs short phrases of text generation quality, but some of us fixed that issue regardless with out crippling the Modeling Core, or toasting multi modality into a static loaf of crap.