Cyril Diagne
cyrildiagne.bsky.social
Cyril Diagne
@cyrildiagne.bsky.social
AI indie hacker. Prev: founded Clipdrop (YCW21, acq. Stability AI), resident at Google A&C Lab, Prof. & head of MID at ECAL
Alright who’s making this
March 7, 2025 at 5:08 PM
How does one keep track? The monodepth/tracking field these days:
December 6, 2024 at 10:05 AM
Align3R: estimates camera poses and consistent depth maps from monocular videos.

Combining it with trackers like Cotracker3 or SAM2 could unlock many fun applications! (cf: VideoDoodles by Yu & al)

Project page (with demo): igl-hkust.github.io/Align3R.gith...
Code: github.com/jiah-cloud/A...
December 6, 2024 at 9:48 AM
Am I the only one amazed that this is what 2*4TB (with thermal case) looks like now?
December 4, 2024 at 9:41 AM
😮 - wall street seems to take the news well though
December 2, 2024 at 4:24 PM
The past few months have been... intense! There's still quite some work to do before the finish line, but excited to launch in the coming weeks 💪⚡️
November 27, 2024 at 4:42 PM
Resemble Enhance seems pretty good: github.com/resemble-ai/...
November 22, 2024 at 8:39 AM
SAMURAI: improve the tracking robustness of SAM2 with 2 main contributions:
- adding motion information to the mask selection
- curating the memory bank based on motion cues

Project: yangchris11.github.io/samurai
Code: github.com/yangchris11/...
Paper: arxiv.org/abs/2411.11922
November 21, 2024 at 8:17 AM
Pyramid Flow is quite impressive for img2video, given than it was only trained on public datasets. Clearly not as dynamic and stable as commercial solutions, but the gap seems to be closing github.com/jy0205/Pyram...
November 20, 2024 at 9:35 AM
A bit surprised with this data from Clerk on sign-in methods preferences: From a sample of 2.5M sign-in, <2% of users chose to use magic links.
November 20, 2024 at 9:31 AM
Yes as a new comer from just a few days I've been a bit surprise by the poor relevance of the Discover suggestions, and lack of alternative algorithms. I guess it's a tricky balance and they've favored a smooth transition from X, which they've nailed. It's been super smooth.
November 18, 2024 at 6:40 PM
Since they don't need the motion direction they just use frame diff which makes it very cheap to compute!
November 18, 2024 at 12:22 PM
Very cool idea for working with videos and transformers: remove duplicate patches and add length embeddings instead. They get a 40% speedup finetuning Vit-L using example packing.

Run-Length Tokenization (NeurIPS 2024 spotlight):
Project: rccchoudhury.github.io/rlt
Code: github.com/rccchoudhury...
November 18, 2024 at 9:42 AM
Indeed!
November 15, 2024 at 6:09 PM