Cohere Labs talk: region-based tokens make long-form video understanding scale with content, not pixels
Cohere_Labs · x · 2026-11-10
Cohere Labs community hosted UIUC PhD student Savya Khosla for a talk, "Scaling Visual Intelligence for High-Resolution and Long-Form Video."
The problem: for a robot or wearable assistant to "watch an entire day" (e.g., help a blind person find keys), current vision systems are prohibitively expensive — fixed patch grids quadruple token count when resolution doubles (even for blank walls), and every frame adds hundreds more tokens even when barely different.
The approach: region-based representations encode each meaningful entity as a single token that persists across frames, so computation scales with content rather than pixels or frame count. The session invites deep technical questions from the Cohere Labs research community.
More from Multimodal
- MiniMax H3 Image lands on Hugging Face with 6-step 4K generation — multimodalart · 2026-11-11
- Adobe launches Firefly Builder to turn ideas into prototypes and working apps — chrisfirst · 2026-11-11
- Turn any song into a vocal-synth cover in 3 steps with Mirelo audio-to-MIDI and Chimera — TomLikesRobots · 2026-11-10
- Fable 5.2 hailed as best image model with Python-code images scoring 1.1M each — thisiskp_ · 2026-11-10
- Pattern Observatory: CGM trend review becomes game analytics mechanic — CurieuxExplorer · 2026-11-10
- Designer builds AAA game concept for Type 1 diabetes with MiniMax H3 Image — CurieuxExplorer · 2026-11-10