@abdallah_dib Nastiest bug I hit: face winding. If your mesh normals point away from the camera, the render looks completely fine. Correct image, correct colors. Zero gradients. Nothing flows.
Everything else? Didn't need a single change.
Now runs on Python 3.11+. Current PyTorch.
This library can reconstruct a full 3D face from a single photo. Mesh, textures, lighting, all of it. And it couldn't even be pip installed anymore. Renderer abandoned, deps broken, pinned to Python 3.8.
Ported the whole rendering backend onto Mitsuba 3.
https://t.co/S3SjnvDpdO
@abdallah_dib It was the gradients. Mitsuba uses DrJit for autodiff. The optimizer runs on PyTorch. Two completely separate autograd systems that don't know about each other.
I wrote a custom bridge bw them. Forward in DrJit, rerender with AD on during backward to pull gradients back into PyT
Stop building UIs for your AI projects.
I made Claude Code the entire frontend for a chess tutoring system. Engine, MCP, spaced repetition all run in the background. Claude just talks to it.
Found more bugs in 2 days than a week of writing tests.
Link: https://t.co/LA17XwWgHI
Opening beta now.
If you work with video editors (or you are one) and this sounds familiar...
We're letting in batches of early users.
Sign up here: https://t.co/asj8WhPv0g
In my company, I work with video editors every day.
Same 2 problems kept killing us:
WhatsApp crushes video quality (good luck downloading on spotty networks)
Timestamp feedback = hell
Tried the "pro" tools. Too complicated.
So we built our own.
Been using it internally for the past month.
Feedback loop went from hours to minutes.
Editors aren't drowning in WhatsApp threads.
We can review on our phone between meetings.
What will it take to convince @googlemaps that the number I'm adding for my business with proof attached, is actually my number I use for that business.