Elephant in the Fridge: Constant-Memory Frame Packing for Long Video Understanding
FramePack-U claims fixed-memory video understanding by packing prompt-relevant, diverse frames into a constant token budget.
The paper says the method works at inference time with off-the-shelf vision-language models, without fine-tuning or architecture changes. It scores frames by semantic similarity to the prompt, ranks for diversity, then packs selected frames at mixed resolutions into a fixed-length context window. In experiments on LongVideoBench, MLVU, and Video-MME, it beats full-resolution baselines and prior token reduction or keyframe selection methods. On Qwen3.5-35B-A3B with a 128-frame budget, the reported gains over the official model are +12.9%, +10.6%, and +9.7% on those benchmarks. HF Daily Papers' note
The paper says the method works at inference time with off-the-shelf vision-language models, without fine-tuning or architecture changes. It scores frames by semantic similarity to the prompt, ranks for diversity, then packs selected frames at mixed resolutions into a fixed-length context window. In experiments on LongVideoBench, MLVU, and Video-MME, it beats full-resolution baselines and prior token reduction or keyframe selection methods. On Qwen3.5-35B-A3B with a 128-frame budget, the reported gains over the official model are +12.9%, +10.6%, and +9.7% on those benchmarks. HF Daily Papers' note
score 6