>>29717734
This is what I think is happening:
You type a prompt and grok takes it and "enhances" it so the model can have something to work and be able to spit out a relatively decent result even if you type nothing or if you just type "ass".
Then it splices that with what it perceives with clanker vision from the first frame and tries to generate a video.
It can fail and get moderated at several points:
- during the prompting, if it thinks it's too nsfw it will fail outright, and there are banned themes (puts it in her mouth, licks X, etc)
- during latent generation, if it scores belows a threshold it will abort (probably to save resources).
- If it passes both those tests it generates a video and does another pass (this is when it fails at 70%+) it checks frames with its clanker vision and if it sees nipple/vag or a full asshole it fails and sends it directly to Elon spankbank with all your information.
This is why nips sometimes get past it in a couple frames and sometimes not, I've had even full vags/clits show up for a couple of frames, it's just luck of the draw.
Now my experience is entirely based on i2v. I don't really do t2v because I got bored of scrolling through blurry pictures, I think t2v doesn't have the prompt check it just generates whatever and blocks it if it sees something on the latent.
What I found is that if we can figure out the weaknesses of the clanker vision it's easier to bypass it, there's a picture I used that for some reason can do a lot of crazy borderline porn stuff, no idea why, maybe how it's converted to latent and back throws the clanker vision off.
