Closed johndpope closed 4 months ago
@johndpope have you tried implementing the paper using gpt4-o ? it looks better than opus
Just playing now on iPhone I did through paper into prompt - it confidently spat out garbage. It might be great to hone in on aligning the diagrams to code (see reference folder - Claude got this wrong had to correct it) - will test later.
Also I did update training code with a face alignment - 2d points - can adjust later 3d - I need to start debugging the training but looks much better than prior work.
UPDATE it seems to truncate file uploaded - can't handle the context window like Claude - 200k waste of time.
Since we predict 3D instead of 2D keypoints, We use an orthographic projection to project the keypoints to the image plane before computing the loss. https://ar5iv.labs.arxiv.org/html/2011.15126
https://github.com/1adrianb/face-alignment