This might be an obvious thing, but have you tried copying the full prompt from perchance’s painted anime images? What I mean is like… Let’s say that I put in the prompt, “A magical girl with a rainbow dress.” But when I save it in my private gallery and look at the prompt, there are some other things added to that:

A magical girl with a rainbow dress., anime masterpiece with soft lighting, gentle intricate style, highly detailed, pixiv, anime art, 4k, art from your name anime, garden of words style art, high quality. Overall, it’s an absolute world-class masterpiece soft beautiful anime-inspired artwork. It’s an aesthetically pleasing, soft, beautiful anime art with impeccable attention to detail and beautiful composition.

(This is for “soft anime,” as that’s my go-to, so the prompt for “painted anime” would be a little different. But the steps to find it would be the same.)

How about sending this to the dev in a more private way (such as through the feedback form on a different generator) so that you’re not linking a bunch of people to it…? 😅 I’m glad you’re trying to get it taken down though.

EDIT: I know that the feedback forms say they’re public, technically, but I’ve never actually seen anywhere it’s publicly posted. At the very least it’s probably much more private than Lemmy.

What device are you using and what browser?

I made a lengthy post about this on the Perchance Reddit in the last couple weeks, but I’m glad you posted this here, because anecdotally, I’ve definitely been seeing an uptick in people (so far all mobile users) saying that they have run into this in the last couple of months and I think the more info we can provide, the better the chances of the dev being able to troubleshoot it. As far as I can find, once it happens the first time, your chats will continue to repeatedly get deleted every so often on that device. Clearing cookies and cache does not seem to fix it. So far, the only fix I’ve seen reported is installing a different browser or reinstalling the browser.

Sometimes I’ve gone quite a while without things getting wiped and I’ll end up with 10-20 chats. Other times, it gets wiped three times in one day despite having only one chat that wasn’t even very long. I haven’t been able to find any sort of pattern. It’s not even about inactivity, because it can get deleted mid-reply (i.e., one AI reply works fine, but then I go to send my reply, and at some point while I was working on my reply, my data got deleted) I’ve decided to take a break from using it on my phone for a bit, until the dev has had a chance to troubleshoot the issue.

@perchance@lemmy.world, if you happen to see this, I’d love to know if it would be helpful for people affected by this bug to submit a feedback form through the website every time they encounter this error, or if that’s not informative and would just be spamming you. 😅

Seconding this. I have a local LLaMA 2 model, and when I asked it (with no system prompt) what model it is, it told me it’s chatGPT-3.5. Then I responded to that message asking it if it was DeepSeek R1 and it told me yes, it’s DeepSeek R1 😂

I hate to have to say it, but unfortunately, afaik that error means that either the data is corrupted or it exceeds your device’s maximum allowed localstorage capacity. The only thing I can think of would be to try opening the exported file and see if you can figure anything out from there. I tested it and you can unzip the export file using 7zip. Then go into the file folder and find the .json file (it should say something like “perchance-characters-export-2025-11-12.json”), and then right click it to open it in a code editor like Notebook++. (If you aren’t able to open it, though, that probably does mean it’s corrupted.)

The first section is all of your character cards. (If you’re like me and tend to upload your character portraits, it’ll have a lot of sections that look like a cat walked across a keyboard as this seems to be how the URL of the uploaded file gets stored for some reason??) Below that, you can find the text of your threads, although actually parsing through them is kind of confusing. But the info should all be there.

If it is a file size issue, then theoretically it might be possible to make the file smaller by removing some of the stuff, but honestly I don’t know how to do that without breaking things. But I hope you are able to at least recover some of it!

Respectfully, I’d like to voice my disagreement with the assertion that the old model was better than the current one. My experience has been that minus temporary (albeit severe at times, yes) oddities that arise from instability with the model still being a work in progress, the new model is better in quite literally every single way from the old model. I say this only because you specifically said here that you want the dev to drop development on the current model and return to the old one – so this statement is less directed at you, and more so that the dev knows that isn’t a universal sentiment.

Again, this is just my experience (I know many other people are having a different experience), and isn’t meant to imply that you “aren’t allowed to complain” or something. But for me personally when I compare the two models:

• I’ve been able to write far more complex stories with the new model than I could with the old model, and much longer ones too. With the old model, every 20 or so replies, complex stories would fall apart completely unless I babysat the memories and summary and edited them constantly. It would forget key plot details (like a character saying “we need to figure out who cursed me” when the opening scene literally was her attacking a mage and getting cursed as a result), hallucinate things that had never happened, mix up characters, and make erroneous assumptions that it presented as fact. The new model has so far been much less prone to these behaviors (though they do sometimes happen).
• I’ve found the new model much better at writing complex characters most of the time, but I have noticed this fluctuate a lot with some of the updates. But the old model typically struggled severely with nuance. For example, in threads with the premade character Yvette, she either hated my character’s guts or trusted her completely (sometimes waffling between the two unpredictably) – there was no slowly building trust. Similarly slowburn romance wasn’t a thing – everything had to be very all-or-nothing or the AI wouldn’t understand what was going on. Which has not been my experience with the new model. I find that the new model does sometimes get really hooked on certain terms and exaggerate them, but the old model would totally ignore character bios at times (e.g., Yvette’s bio explicitly says she is NEUTRAL about magic, but in so many forbidden romance stories where my character was a witch/mage, it would have Yvette either hate magic or else be completely appalled at the oppression of mages and be very pro-mage – once it literally even had her express genuine shock that some people would hunt mages to cull their mana lmao…).
• Relatedly, the old model often seemed to struggle to keep characters coherent across situations. For example, I once did a Yvette isekai thread and she suddenly went from “what the fuck is this shit” type dialogue to “Ah… I am not certain of what you speak, but would those perhaps be the flying metal boxes I had seen when I arrived? I did not recognize them, and found them quite frightening.” 💀 I haven’t had this problem with the new model.
• With the old model, I felt like I was constantly fighting it to write anything that wasn’t all sunshine and rainbows. It didn’t even have to be a taboo subject – the second any emotional weight appeared, it was constantly trying to change the subject. It also insisted every single character be a paragon of healthy coping skills and never ever handle their feelings in an unhealthy way, and I don’t even mean dangerous things, but even having a character verbally snap at a loved one or hide their feelings, there were times I would explicitly prompt it to do that and it would go, “She thought about [thing I instructed], but realized that wouldn’t solve anything, and decided to [use a healthy strategy] instead.” Also, if a character had a “taboo” thing mentioned in their backstory (like a history of unhealthy behaviors or childhood abuse), it would often bend over backwards to avoid referencing that in any way, shape, or form, which really detracted from the character depth because that part of their history never informed their perceptions, they never even thought about it unless my character directly asked about a related subject (and even then it would try to find any excuse to change the subject). I’ve genuinely never had a problem with the new model going dark when I didn’t intend for it to do so, and in situations where the darkness fits the plot, I love its willingness to go there without me having to metaphorically drag it kicking and screaming. This was my #1 frustration with the old model, and I was pretty close to giving up on Perchance and switching to SillyTavern because of it before the new model was rolled out.
• The new model is much smarter in its ability to infer things and understand implications (though this varies due to the current instability). I’ve been genuinely impressed with it sometimes picking up on foreshadowing that I had thought to be very subtle. Similar to the point above about everything being all-or-nothing, I frequently had to spell things out for the old model very directly before it would figure out what I was getting at. I literally had a time where I wrote that my character “leaned in, her eyes on Yvette’s lips in a silent request for permission” during a heartfelt moment and it literally didn’t acknowledge at all that she was trying to kiss her. Finally in exasperation I edited it to clarify, “in a silent request for permission to kiss her,” and immediately Yvette kissed my character… implying that the AI likely hadn’t understood what that “silent request” actually was, even though that sort of description is pretty common in kiss scenes.
• When writing two women flirting, the old model used to sometimes seem to forget that they were both women. For example the AI character putting her hand on my character’s face and feeling stubble. Or once the character made a flirty double-entendre about something getting “hard,” which… isn’t really a typical double-entendre between two women lol. I’ve never once had that happen with the new model.
• Another thing that makes the new model feel so much more immersive to me is the way that it actually references my character’s cultural background. For example, the user character I write with most is named “Carmen” (not my real name ofc) and the old model would usually shorten it to “Carm” which I hate the sound of so I was always fighting it to stop doing that (editing it out to say “Carmen” only for it to call her Carm again a few messages later). But just recently the new model had a loved one call her “Carmencita” which I loved as that respects Carmen’s cultural heritage, which the old model ignored 100% of the time. I also really love how the new model will have her loved ones call her Spanish-language terms of endearment like “querida.”
• In writing child characters, I often felt like the old model treated them like adults. Telling them things that an adult with healthy boundaries wouldn’t tell a child, using vocabulary or concepts a child that age couldn’t be reasonably expected to understand, placing expectations on them that are wildly age-inappropriate, acting like it’s surprising for a 6 year old to react childishly to something, etc. Whereas I haven’t run into those things with the new model.

I genuinely don’t know why experiences are varying so deeply in terms of people’s perception of the old vs new models. Maybe it’s our writing styles being a better fit for one model than the other or something, or a difference in custom writing instructions, I don’t know. (I don’t think it’s type of story, because I’ve written so many different types of stories with the new model – drama, slice of life, romance, comedy, angst, found family, iyashikei [healing], isekai, fluff, hurt/comfort, smut, adventure, fake dating… – and it’s done beautifully with matching my tone in all of them.) But when I see people advocate for a rollback to the old model, I always feel like I want to let the dev know that not everybody sees it that way, because I found the old model extremely frustrating to write with. (Again that doesn’t mean I think you aren’t allowed to complain. It just means that I also want to share my differing experience.)

Thank you so much for the update and for all your hard work! I’m grateful to know that you are reading user feedback.

By and large, I’ve been really happy with the new text gen. When it works well, it works really, really well. There have been a few times that it was working amazingly and I was really impressed by the quality – times where it took a really complex scenario and performed beautifully, picked up on foreshadowing, did a great job at portraying nuanced characters and writing realistic in-character dialogue, etc. with minimal corrections needed.

I also already find it a huge improvement from the old model in almost every single way, outside of times the model starts severely and repeatedly hallucinating. More than anything else, I’m so, so glad that it isn’t afraid of dark themes like the old model was. I love that if I give a character a dark backstory, it will sometimes add to and embellish it, and will also bring those details up when relevant of its own accord (without me having to repeatedly instruct and nag it to do so). So I’ve been really happy with the update!

I’ve been having this issue as well. I think a recent tweak has caused it to have trouble decoding its context (things it should remember from the character card, summary, previous messages, etc.). It feels like it pulls out a few keywords and hallucinates everything else, including the actual situation where those keywords came up, even when it’s written in the summary and/or character card. Maybe the dev is working on expanding the context window (the amount the AI can remember), and that’s causing some glitches in the meantime?

There are some workarounds to reduce the impact until the dev has a chance to sort it out though. In addition to what others have already said, another thing you can do is edit the reply and use auto-complete. To do that, double-click on the AI message. In a case like this where no part of the reply made sense in-context, I would just delete it all and write something like, “Mona paused, thinking about their beloved children – Felix, and the twins Takara and Aiko still in her belly.” Then you hit the “▶️ auto-complete” button in the bottom right corner of that message, and it should finish the rest coherently. Also if you start the AI’s message in German, I would expect it to finish the message in German too.