Post

norvid_studies
@norvid-studies.bsky.social

why don't more people generate exploratory images with image models? very impressionistically maybe 1 in 5 or 1 in 6 do, among moot cluster selected for broad AI attention/proficiency? share your negative stories

Asa ⭔
@asa.engineer

idk, I think more people should. the output gets interesting when you chop most of the bytes off the model

Asa ⭔@asa.engineer

⭔⬡

2 reposts36 likes

Techy Ben
@techyben.bsky.social

This was my experiment breaking local models (with only words/tokens, not with any changes to the image gen model, just finding latent spaces it could display that were not intended).

Asa ⭔
@asa.engineer

I just use a very small quantization, usually a Q2_K or Q4_0, for a lofi diffusion look and more texture and distortion.

Asa ⭔
@asa.engineer

the model used is z-image base Q4_0 (uniform quantization). The prompt used for most of these is fairly minimal, something like: "a surrealist landscape. the horizon and sky fade together in a smooth gradient. the end of a Lofstrom loop rises from the ground."

Asa ⭔
@asa.engineer

"Lofstrom loop" because I was fascinated with the concept, and it steers to a sci-fi/future/space travel region in latent space, but the tiny model can't possibly depict it believably (good, accuracy is explicitly not the goal) so you get weird, unpredictable hoops, arches, and tendrils.