Spaces:
Running
on
CPU Upgrade
Censorship?
Experimenting, I have found that certain requests just make the whole thing hang up.
I think this is a sort of "censorship", that the system is refusing to draw what I am asking for, because I asked for something inappropriate.
However, sometimes the same inappropriate request DOES work if I change the style.
Perhaps you can share some examples of the prompts you are using?
That would help answer your question.
In other news, it's not clear why you would want to use inappropriate requests unless you are just testing the censorshiping system.
Regardless, I'd say you might be getting good feedback from the system either way.
I do think some hangs are unrelated to anything we enter as prompts.
The Comic Factory space is trending and popular and that seem much more likely to cause delay.
Added: In general, it is my understanding that all online AI generators 'censor' inputs based on some criteria.
It may be to focus the model for specific use cases or it may be to prevent abuse of the system, etc.
I'd consider that much more mundane than actual 'censorship' however.
A better word might simply be 'filtering'.
I have been experimenting more, and it now seems that what makes it "hang up" is rather random, and not connected to the appropriateness or inappropriateness of the prompt. I think sometimes it just does not understand the prompt. Sometimes I can change this by changing a word, or even just changing the spelling of a word.
Hi, indeed sometimes the LLM (llama-2 running on TGI) can fail, I think that's because the result to a given prompt is cached, even when the result is.. empty!
and for SDXL, well the AI Comic Factory is as censored or uncensored that SDXL 1.0 can be, there is no improvement or degradation regarding that
recently, I've started working on a real censorship mechanism for another work in progress project, Panoremix (to prevent from getting banned by Replicate), but I haven't added it to the AI Comic Factory yet (I didn't really need to, since all user generated content is private and ephemereal, except what people decides to share publicly and in their own name, on their HF or social network account)
note: another factor for apparent censorship could be that your prompt is too long, and some concepts or words have been cropped
Can you let us know what the maximum length of a prompt is?
currently I have enforced a size limit of 77 characters (it's in the source code, in main.ts)
the reason is that there is a total limit of about ~380 character limit for SDXL, which needs to contains:
- this user-defined prompt
- the style prompt ("comic book.. vintage.." etc)
- the LLM-generated prompt, which can be a bit long
-> the number could be set to 80, 90, 100.. but this increases the risk that some core prompt components will be cropped, degrading the quality of the story or style
-> if SDXL 2.0 etc or another future more comes one day, the limit could be increased
-> if a LoRA is used one day, the limit could be increased (we would not need to write so many text to define the style)