Video2LoRA: Parametric Video Internalization for Vision-Language Models
Quick Answer
Video2LoRA introduces a novel method for parametric video internalization in vision-language models, enabling SmolVLM2 to answer queries with zero visual tokens.
Quick Take
It reduces visual-token load by up to 1,500x and query TTFT by 6-80x while maintaining performance across multiple benchmarks.
Key Points
- Video2LoRA generates LoRA adapters directly from video in a single forward pass.
- Achieves equivalent performance to direct video-in-context inference across five captioning benchmarks.
- Reduces answer-time visual-token load by up to 1,500x and query TTFT by 6-80x.
- Stable performance up to 1,024 frames and 1024px resolution.
- Supports independent adapter generation for non-overlapping video segments.
Paper Resources
Article Content
From source RSS / original summaryarXiv:2606. 04351v1 Announce Type: new Abstract: Processing video in is expensive: each frame occupies hundreds of tokens, and inference cost scales with every frame and every repeated query. We introduce Video2LoRA, a method for parametric video internalization. A perceiver hypernetwork reads the intermediate representations produced layer-by-layer as a frozen VLM encodes a video, and generates a Low-Rank Adaptation (LoRA) adapter in a single forward pass.
Unlike standard LoRA fine-tuning, which requires iterative gradient updates, Video2LoRA predicts these weights directly from the video. Trained for SmolVLM2 500M and 2. 2B on video summarization and captioning, Video2LoRA enables the same frozen VLM to answer queries from the adapter alone, with zero visual tokens in its context at query time.
Video2LoRA is statistically non-inferior and equivalent to direct video-in-context inference across all five captioning benchmarks at both model scales, and across seven of eight video question answering benchmark-scale pairings. Although trained only on 12 frames at 384px, it remains stable up to 1,024 frames and 1024px, where direct video-in-context inference often degenerates.
Across this sweep, it reduces answer-time visual-token load by up to 1,500x and query TTFT by 6-80x, while preserving video-faithful outputs. We also find that independently generated adapters for non-overlapping video segments can compose in rank space, suggesting a path toward chunked long-video internalization.
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CV
See more →ProMoE-FL: Prototype-conditioned Mixture of Experts for Multimodal Federated Learning with Missing Modalities
ProMoE-FL introduces a Prototype-conditioned Mixture-of-Experts framework for multimodal federated learning, effectively addressing missing modalities. It outperforms existing methods on four chest X-ray datasets, demonstrating superior feature synthesis capabilities in both homogeneous and heterogeneous settings.