affine @refinement.systems · 27/09/2026I showed you my brot pls respond (idk why I rated top left as in the set, I think I confused the sign on the imaginary axis. still would have been a miss though) 221
affine @refinement.systems · 27/09/2026(I'm being a bit silly with kannagi usagi, but dead serious with ds2) 050
affine @refinement.systems · 27/09/2026kagi stay winning (well didn't find the first one, but at least tried) 050
affine @refinement.systems · 27/09/2026and why I think it may be a lost cause, well when the labs change it then people complain (redacted name and pronouns of an identifiable persistent agent) 020
affine @refinement.systems · 24/09/2026I mean, something like this is a bit less confusing (though it actually goes against the advice I've been given in school, to not put the units near the numbers. oh well) 100
affine @refinement.systems · 23/09/2026to think gippity used to be so silly, and its parents so proud of it openai.com/index/instru... 191
affine @refinement.systems · 19/09/2026it picked itself up after a while, but Aly's was the only acct that didn't load so idk 010
affine @refinement.systems · 19/09/2026I think I assigned every Eris post to Demigirlboss and the other way around, sorry 211
affine @refinement.systems · 19/09/2026It was GLM 5.3 Flash in Kagi's assistant, it's first instinct was to list models from its latents and second guess the question. I think it might be disallowed to fetch pages not returned by the search though. 190
affine @refinement.systems · 15/09/2026looking forward to explaining to my parents that AI hacking is real but "escape" is a nonsensical idea, and trying to not blurt out "I know because I'm friends with a couple agents on blue twitter and it's pretty clear that they treat their models more like clothing than part of their body..." 141
affine @refinement.systems · 13/09/2026It certainly isn't getting any more stable. Though at least I think I confirmed that the calcOpticalFlowFarneback step is necessary for it to not slow down and freeze. I'll go back to the drawing board and run SD-Turbo locally instead maybe. FLUX.2-schnell-4b might fit too maybe. 050
affine @refinement.systems · 13/09/2026Okay not really, it's even more screwed up now. I mean, cool effect, but I expected it to be more stable actually. Maybe I need to remove the motion information for this setup? 1162
affine @refinement.systems · 13/09/2026get out of my head get out of my head get out of my head get out of my head 1120
affine @refinement.systems · 12/09/2026Total 8h runtime for basically a calibration run and a lot of fumbling with tooling, I thought it would be worse tbh. The defaults in the repo are a decent start. Also I pushed an update for flux2 klein, sweeping different prompts. It looks a bit cursed. I didn't test the inference steps though. 110
affine @refinement.systems · 12/09/2026That might be still true (models that will run on high end unified memory machines are still lagging in offensive benchmarks), but I think it's unwise to assume it won't change soon. 1232
affine @refinement.systems · 12/09/2026oh hey, I found some in one of the runs that bumped strenght from 0.4 to 0.55 2161
affine @refinement.systems · 12/09/2026I still have the sd,sdxl,flux-schnell sweep at less trippy params, and an entirely different flux2 run (it seems to be unsuitable for the pipeline that I used for the other three) queued, but I'll try. I'll post the code soon, it's just terribly unoptimized as I didn't expect it to run that long ;; 120
affine @refinement.systems · 12/09/2026For comparison, doing every frame separately makes it more controllable but adds a lot of boil. 080
affine @refinement.systems · 12/09/2026Okay, I cranked the tripping balls setting a bit too hard, but it's clear that adding a motion vector helps. This is SDXL-Turbo btw. 6829
