It uses Chroma Flux, one of the slightly lower quality/older models. I'm trying to figure what exact version it is. No, it's not the Chroma HD, but I might be Flash also

No, you are good! I understand that, it does seem that way tbh Not everything is free and open source. One software can be proprietary and free and the other can be open source and paid. True, a lot of AI models aren't open source, but we are lucky to have a few. I am not sure why I though DeepSeek, probably because it was free on the chat ui. Going through a whole reverse engineering would be complicated for me LOL. I think I knew they published like the weights and code but I was not aware the dataset was closed.

You were right all along! It is a Google model, more likely Gemma 4, or another Gemma model, or some other Google model idk lol like you mentioned. I used a few AI text generators as examples and both confirmed to be language models trained by Google! So know we know the AI text plugin is using a Google-trained AI text model. I wish we knew the exact model

I think only that AI text generator was using DeepSeek through huggingface api, not the actual ai text plugin. I didn't mean to mistake DeepSeek as the default text model for Perchance, I agree with you, it does seem to be more likely Gemma for sure.

I found this in Elements tab from DevTools:

• "the text model is likely gemini model gemma4" (2d ago)

from https://perchance.org/ai-text-plugin

I didn't see it on the page, it was hidden. I don't know if it was from a comment or message (there was comment/message section) but apparently (if its an actual person who written it and not hidden text in code LOL) else believed it was also using Gemma 4. The other AI text generator I must've been using more likely used an external source (Huggingface and Deepseek, according to Network, seem like strong indicators)

And the fact it said the cut off date was Jan 2025 and it kept insisting it was Google makes me believe Perchance is actually using Gemma 4 or other Gemma models.

You may be right about it being Gemma 4! And it might be possible for Perchance to allow multiple AI text models, whether through the plugin or custom coding (idk lol). I have seen some text generators using supposudly using different models from apis but I don't know if its actually true or not. I suspect that https://perchance.org/ai-character-chat might not actually use Perchance's text model plugin. Based off of what you asked/instructed/prompted, it may be very well possible its Gemma 4, especially Google releases their AI text models as free and open source, so I'd least be a little surprised that's what the ai text plugin was using LOL

And I am even more convinced you might be correct because I found this on Google's website:

Earlier versions of generative models were only able to process 8,000 tokens at a time. Newer models pushed this further by accepting 32,000 or even 128,000 tokens. Gemini is the first model capable of accepting 1 million tokens.

https://ai.google.dev/gemini-api/docs/long-context

Which sounds similar too what your AI said

I actually found out through using DevTools to extract/analyze the data! Using network, I found these:

1. https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/resolve/main/tokenizer.json

2. https://huggingface.co/api/resolve-cache/models/deepseek-ai/DeepSeek-R1-0528/4236a6af538feda4548eca9ab308586007567f52/tokenizer.json?%2Fdeepseek-ai%2FDeepSeek-R1-0528%2Fresolve%2Fmain%2Ftokenizer.json=&etag=%22d81c3a6e2093a07ae6fa596d9b154f6669aafafd%22

3. https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/resolve/main/tokenizer_config.json

4. https://huggingface.co/api/resolve-cache/models/deepseek-ai/DeepSeek-R1-0528/4236a6af538feda4548eca9ab308586007567f52/tokenizer_config.json?%2Fdeepseek-ai%2FDeepSeek-R1-0528%2Fresolve%2Fmain%2Ftokenizer_config.json=&etag=%22c24fd8eb5753a737488f8e447cb5739b0e9f7a1a%22

You're right, there was no DeepSeek model api, I think I meant to say the huggingface api.

Only from https://perchance.org/ai-character-chat though. I haven't looked at other AI text generators/tools.

Ah, okay. I see horde-safety kinda like integrates the Fedi code in their filters.

Fedi-/Horde-Safety use an image interrogator, to query what’s actually in the image. That’s immune to that. They also have all these heuristics. (You can have a look in their code.) It’ll come with false-positives, though. It’s a trade-off.

Huh, now that's pretty interesting. I will defiantly take a look into their code, I want to see how they programmed it.

Mind you’re currently in the process of fleshing out your own interpretation of the wheel. My recommendation was not to do that. And your approach has the usual downsides: People can use other words, synonyms. Describe it. Or leverage wrong spellings. Jam an adjective in-between. Do an exact request, but in Chinese language… All that will circumvent filtering by simple word filters. And that’s usually what happens 5 minutes after you roll out such a system. It’s a) pretty much not effective at all. And b) it’s complicated and you’ll end up in a cat and mouse game.

Yeeaaaahhhh...I just realized that that now...emphasis on the process of fleshing out my own interpretation. I wouldn't recommend as something to actually make and integrate as censorship. I think I was literally doing that, trying to showcase how I thought how the censorship/filtering worked, like the basics of it...I knew it wouldn't be as accurate or effective because not all censorship/filters actually work like that...I guess in my head it was more of a "demonstration" rather than an actual tool/solution. I think I meant to demonstrate how the censorship/filter on what words and associations to focus on instead of banning innocent prompts because of too much censorship (Unfortunatly, that what Grok does. Because they effed up knowing releasing the ability someone's post without consent and now it all backfired, cranking up the censorship and making it practically unusable. And the fact the government is enforcing it also makes it worse) ...But you are 100% right: its not effective since people can use ways to circumvent the filter like you said, use different words, synonyms, mispellings, sentences, etc. (Man, I wouldn't be too surprised if it was less than 5 minutes...its insane...). Very ineffective and complicated...I just thought about its trying to keep adding words when people can bypass is hard as hell. And the cat and mouse game? God, couldn't be more true. Trying to add more words to the filter and people STILL find ways to bypass would be exhausting.

I generated these with my prompt. I think this is what you were going for. Some of them were "cherry picked" as my way of prompting doesn't always work, but these were absolutely generated with my prompt, many of them actually (when I increased the numbers of images). I used some of the AI image model's (a random generator from Perchance) settings like guidance and art style (didn't really help much until I put real human being/photograph in the prompt...bruh moment). No flyaways, no buns :3

I can share a erm vanilla version of the prompt, should get the about the same result:

woman with long and tall ponytail that sits up high on top of head, the ponytail is pulled-back-tightly/slicked back with gel and stretches the forehead. The woman is a real human being. She is in a professional photograph. Avoid styles, increase guidance (try 12-15), and increase number of images (10 is okay, 20 is good, 30 is great, 40-60 is amazing!)

I'm thinking like (for programming filter only):

contains/includes/with = (words go here) exclude/omit = (words go here)

Now, you can't include everything at first, just keep adding and refining the filters over time or from reports from users. Very minor idea:

Words (adjective, noun): little, young, small, tiny Containing/including with (nouns): boy, girl, child, kid

(can even add suffix/prefix for like pre-teen, pre-adolescent, etc)

Exclude/omit from (adjective, verb, adverb, preposition): [you get the idea]

if there's a word with (-s, -es, -ed) would also include that too...

This tells the filter to what to scan or analyze for. Whether the words are categorized or not (doesn't matter) the core concept remains the same. I think there may be similar pre-made solutions or tools on github like my idea, I would also look into those. Its not as hard as you think, just a bit time consuming to configure...