Post

I'm trying to replicate the DLSS5 haywire video with open models, no shroom gnomes yet but fun results so far (200 iterations of FLUX.1-schnell and a prompt "Computer, enhance!")

10 reposts86 likes

Same with SD-Turbo. I'm literally just running torch in a loop with the model, I didn't try replicating normal DLSS operation (I think it's supposed to be `processed_n = model(processed_{n-1}, fresh_n)` or something like that?)

I’m unhappy with my results in this space thus far (since 21 actually, shocked pikachu) but I’m convinced there’s something here. you might hear from me on this point again one day.. (these are great, not quite what I’m going for though)

nah, naively blending the previous previous processed frame with a fresh one is too dumbed down, looks like crap when running it on a video.

Could be interesting. Anyway the code is public: bsky.app/profile/did:...

affine@refinement.systems

Ok, I cleaned up the code a bit, but idk if it works after the last minute reorganization into separate scripts: github.com/refinement-s... In any case it currently needs ~33+ GB VRAM to run with all models currently tested, the offload may bring it down to fit in 24, but this was ~10x slower

how many seconds per second are you averaging? i got some interesting results (at real time on camera input) with metalfx but i don't have kernels trained on text and comics there, so they aren't generative; more like the second thing