Good choice ICML, congrats 2014 Berkeley gang!

3DV has asked me to ask you to ask me anything (presumably about 3D vision, but I'm flexible)

diffusion + layering is going to be so so useful

oooh, compelling demo …ian-scale-space-fields.mpi-inf.mpg.de

Easy way to fix academia: Add a negative offset to all citation counts in Google Scholar. Every new paper starts at -100 citations and has to claw its way out of that hole.

Whoa looks like there's gonna be a Starline telepresence demo at SIGGRAPH next week s2024.conference-program.org/presentation/?… I haven't demoed Starline in a while but all of my early demos were extremely impressive, comparable to the feeling of trying VR for the first time.

SIGGRAPH award for SMERF! Gonna keep stacking letters in front of "RF" until it stops working. smerf-3d.github.io

lpanaf.github.io/eccv24_glomap/ seems really good? "In terms of accuracy and robustness, we achieve results on-par or superior to COLMAP, the most widely used incremental SfM, while being orders of magnitude faster."

A random walk(*) through #flux latent space

Playing around with Flux+LoRA, very fun. These results are from fal.

I've observed a high correlation between train time and identity preservation in Flux LoRA/DreamBooth services. People seem to be undertraining their models to save compute for generation, which is likely the wrong call. Better to have a small number of high quality generations.

Quick and easy ~3D Dreambooth experiment: I trained a LoRA (via astria.ai, highly recommended), prompted "a character reference sheet of JONBARRON in a Star Trek cartoon", cropped+centered two faces, and Dream Machine'ed them. Tada, surprisingly stable 3D rotation.

The year is 2040, and you are playing what is considered to be a typical 2040 video game. The game engine (which may be generative) that renders pixels, where is it running? And the engine that powers dialogue and character AI, where is that running?

Has anyone crunched the numbers on the watts-per-pixel performance of gamengen.github.io compared to just running Doom the usual way? I assume it's somewhere in the 10-20 OOM range but I really don't have a clue.

Am I crazy or does the new A$AP Rocky video have AI art aesthetics? It feels they took some good text-to-image slop as a reference and then lovingly recreated it with practical effects. youtube.com/watch?v=5URefV…