Spaces:
Running
on
CPU Upgrade
the voice falls apart after 3 generations
This is clearly one of the most interesting projects! I just found it yesterday. But have you noticed that the "voice" falls apart hard after about three generations?
Take the prompt both as a prompt in its own right and a stand-in for the sense of limitless possibilities: something weird but that I could show my parents.
After about 3 rounds of a mensch and a floating legal courier, you get a very stubborn and persistent gravitational pull towards the boring, tedious and banal. The potential for weirdness drains away very hard and it genericizes into:
cobblestone streets
market stalls
a recurring affluent redhead
a corporate ethos
managerial types in business suits
stock footage for corporate videos
groups of affluent twentysomethings
captive-audience meetings
lobotomies
advertising
marketing
root canals
Pieces of the original fall away and it gets more and more drab. It is just nothing to do with it anymore.
The issue is not the hallucination - levitating ducky couriers are the possibility of a new kind of psychedelic film, or a new kind of thing, which you, Julian, are closer to than most people on here. I would like the hallucination to sustain please!
It probably comes down to money, like the aggressive switch to Zero. Fine, but isn't it just as expensive to render the dreadful boilerplate over and over?
thank you
whichever
Hello no i haven t see it but so many thing to look at ! happy to have some news about you !
Hello!
the "voice" falls apart hard after about three generations
what do you mean by voice, do you mean that the story/content doesn't evolve in a good or interesting way?
if that is the case, then it is most likely a "bug" in the prompt I created to make this evolution (this is the first version of the loop, so there are still things to fix and improve).
This is a beta feature so if it doesn't work well enough, you can disable it in the settings and then get more static videos (eg. for a fishtank.
Note that the content/prompt quality issue can also be caused by the LLM engine: by default Tikslop uses a very low quality LLM that can often produce weird or incoherent results, especially after a few generations. Tto get a better prompt evolution quality, you can switch the LLM in the settings (but a HF token is required)
Overall, I would say this whole project is still experimental, and progress is slow as I can only work on Tikslop from time to time (to make Tikslop better, I need to make a lot of iterations and mistakes)