I agree - setting up an environment is so painful and while claude/gpt/etc are okay at helping dependency issues if codex can automate this it will be so nice

looks very cool! am excited to compare this to megasam

okay I fell for this and now I’m saving this monumental piece of work 😭😭😭☠️☠️☠️

ahh this is my mistake - it is just image, pointmap datapoints

I am still impressed by sora’s ability to learn 3D without 3d info (like pointmaps)!

i’m not gonna lie the naming is hilarious 😭😭☠️ alligator va croco

post submission cycling is like mcdonalds sprite after a hangover

I’m honestly still shocked at this - how did Sora manage to learn 3D consistency just from videos?

Dust3r (i know it’s completely different) had to be given camera parameters during its training process, so to me it makes sense it was able to understand 3D very well

if you have a CUDA machine, you can try nerfstudio’s splatfacto method if you want a video of the optimization process on custom data

there are some good videos on the nerfstudio documentation website too!

docs.nerf.studio/nerfology/me...

the way cameraHMR managed to get the hands flat for the handstand picture was very neat!!