LLaVA-OneVision-2.5 replaces frame sampling with codec signals, wins 13 of 17 video benchmarks
jiqizhixin · x · 2026-11-19
- LLaVA-OneVision-2.5, led by GlintLabAI and lmmslab, lets video models learn from the codec instead of sampled frames — and it finally beats frame sampling on general video tasks.
- The route took three steps: OneVision-Encoder showed a vision encoder can align with codec signals; LLaVA-OneVision-2 gave first evidence inside a full VLM; now 2.5 delivers the win.
- Core mechanism: instead of 1 frame per second, the model uses the codec's bit-cost to locate regions carrying new information and packs them into a few canvases.
- The same 4B weights run in two modes: offline for images and long videos, and online streaming where the model decides at every step whether to speak or stay silent.
- Results: across 17 video benchmarks, codec input beats the model's own frame sampling on 13 and Qwen3.5-4B on 16, with a unique full-frame-rate streaming design.
Related event: LLaVA-OneVision-2.5: Codec Input Beats Uniform Frame Sampling(5 posts)→
More from Models
- Anthropic allegedly profiles users and silently downgrades them to weaker models at full price — JafarNajafov · 2026-11-19
- Antigravity CLI 1.5.12 screenshot shows Gemini 4 Neon (Medium) model option — patloeber · 2026-11-19
- Subscription value test: Gemini ~26x vs OpenAI ~10.5x API credits; real-bug eval underway — PawelHuryn · 2026-11-19
- Daily digest: Xiaomi MiMo spinoff rumor, Tencent cuts Yuanbao spend, Gemini 4 Neon ships — 快鲤鱼 · 2026-11-19
- Branching a ChatGPT workspace wipes the agent's context, user complains — LucaAmb · 2026-11-19
- 0.5B ASR model hits 6.96% multilingual WER, beating Meta's 15x-larger omniASR — thisguyknowsai · 2026-11-19