We released OCTOPUS 🐙
KV-cache compression using octahedral parametrization: rotate keys, group them into 3D triplets, encode each direction, and decode directly inside attention.
Paper: arxiv.org/abs/2605.21226
Project Page: octopus-quant.github.io
I’m the Co-Head of 3D & Image at Stability AI with research interests in the intersection of machine learning and computer graphics
- Inspired by the excellent tev (github.com/Tom94/tev) I made a version for videos - especially with split views (Compare two videos), error vis, deep zooms (200x), HDR support. github.com/vork/Framewise
- Current hobby: vibecoding tiny tools that eliminate annoying tasks. I open-sourced them in case they’re useful to others.
- We’re hiring 3 researchers for the Stability AI 3D team — and with our new EA partnership, this is an absolutely massive opportunity. If you’re passionate about 3D, graphics, AI models, or VLMs consider applying below:
- If you’ve been affected by the recent Meta layoffs and your work touches 3D, video, or image generation - please feel free to DM me. We’re hiring at @StabilityAI.

