Frames2LoRA: Parametric Video Internalization for Vision-Language Models

• URL: http://arxiv.org/abs/2606.04351v2
• Date: Wed, 10 Jun 2026 09:50:06 GMT
• Title: Frames2LoRA: Parametric Video Internalization for Vision-Language Models
• Abstract summary: Frames2LoRA is a method for parametric video internalization.<n>A perceiver hypernetwork reads the intermediate representations produced layer-by-layer as a frozen VLM encodes a video.<n>It generates a Low-Rank Adaptation (LoRA) adapter in a single forward pass.
• Score: 52.46510577817688
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Processing video in vision-language models is expensive: each frame occupies hundreds of tokens, and inference cost scales with every frame and every repeated query. We introduce Frames2LoRA, 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, Frames2LoRA predicts these weights directly from the video. Trained for SmolVLM2 500M and 2.2B on video summarization and captioning, Frames2LoRA enables the same frozen VLM to answer queries from the adapter alone, with zero visual tokens in its context at query time. Frames2LoRA 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.

Related papers

• TAME: Temporal-Aware Mixture-of-Experts for Text-Video Retrieval [1.4874449172133888]
  Text-Video Retrieval (TVR) retrieves videos that match a natural-language query.<n>TVR is limited by the lack of temporal modeling.<n>We propose Temporal-Aware Mixture-of-Experts for Text-Video Retrieval (TAME)
  arXiv  Detail & Related papers  (2026-09-02T07:12:50Z)
• Don't Pause: Streaming Video-Language Synchrony for Online Video Understanding [69.296913137409]
  We introduce a novel paradigm for online video understanding: Streaming Video-Language Synchrony (SVLS)<n>We present LyraV, a live streaming assistant built upon a hierarchical control framework with two core innovations.<n>First, the Frame-Driven Transition Controller (FDTC), makes high-level semantic decisions on when to continue speaking, start a new response, or stay silent.<n>Second, the Streaming Token Pacer (SToP), a plug-and-play lightweight predictive module, dynamically adapts the language generation rate to match the pace of the visual content.
  arXiv  Detail & Related papers  (2026-06-05T07:29:20Z)
• AdaCodec: A Predictive Visual Code for Video MLLMs [66.30108093864139]
  We call this interface a emphpredictive visual code, and instantiate it for video MLLMs as textbfAdaCodec.<n>AdaCodec spends full visual tokens on a reference frame only when its conditional predictive cost is high.<n>It encodes inter-frame changes, including motion and prediction residuals, as compact P-tokens.
  arXiv  Detail & Related papers  (2026-06-01T17:56:35Z)
• DynFrame: Adaptive Reasoning-Driven Multimodal Framework with Dynamic Frame Augmentation for Complex Video Understanding [19.492762364136393]
  We present DynFrame, a framework that emits the temporal window and the sampling density as native tokens within a single autoregressive pass.<n>This learnable span-density retrieval enables acquiring multi-granularity evidence with a single retrieval step.
  arXiv  Detail & Related papers  (2026-05-26T08:16:16Z)
• Towards Effective and Efficient Long Video Understanding of Multimodal Large Language Models via One-shot Clip Retrieval [57.88666884515147]
  We propose One-shot video-Clip based Retrieval AuGmentation (OneClip-RAG)<n>OneClip-RAG makes full use of the merits of video clips for augmented video understanding.<n>It is also equipped with a novel query-guided video chunking algorithm.
  arXiv  Detail & Related papers  (2025-12-09T09:40:20Z)
• video-SALMONN S: Streaming Audio-Visual LLMs Beyond Length Limits via Memory [51.03819128505358]
  Video-SALMONN S is first to process 3-hour videos at 1 FPS and 360p resolution under a fixed memory budget.<n>A test-time-training memory module continually updates token representations to capture long-range dependencies.<n>A prompt-dependent memory reader retrieves context-relevant content from fixed-size memory.
  arXiv  Detail & Related papers  (2025-10-13T08:20:15Z)
• Temporal Chain of Thought: Long-Video Understanding by Thinking in Frames [70.93346841539626]
  We present Temporal Chain of Thought, an inference strategy for video question-answering.<n>We use the VLM itself to iteratively identify and extract the most relevant frames from the video.<n>We demonstrate how leveraging more computation at inference-time to select the most relevant context leads to improvements in accuracy.
  arXiv  Detail & Related papers  (2025-07-01T18:39:26Z)
• Chain-of-Frames: Advancing Video Understanding in Multimodal LLMs via Frame-Aware Reasoning [37.86612817818566]
  We propose to obtain video LLMs whose reasoning steps are grounded in, and explicitly refer to, the relevant video frames.<n>Our approach is simple and self-contained, and, unlike existing approaches for video CoT, does not require auxiliary networks to select or caption relevant frames.<n>This, in turn, leads to improved performance across multiple video understanding benchmarks.
  arXiv  Detail & Related papers  (2025-05-31T00:08:21Z)
• Multimodal Long Video Modeling Based on Temporal Dynamic Context [13.979661295432964]
  We propose a dynamic long video encoding method utilizing the temporal relationship between frames, named Temporal Dynamic Context (TDC)<n>We segment the video into semantically consistent scenes based on inter-frame similarities, then encode each frame into tokens using visual-audio encoders.<n>To handle extremely long videos, we propose a training-free chain-of-thought strategy that progressively extracts answers from multiple video segments.
  arXiv  Detail & Related papers  (2025-04-14T17:34:06Z)
• VidCtx: Context-aware Video Question Answering with Image Models [15.1350316858766]
  We introduce VidCtx, a novel training-free VideoQA framework which integrates both visual information from input frames and textual descriptions of others frames.<n>Experiments show that VidCtx achieves competitive performance among approaches that rely on open models.
  arXiv  Detail & Related papers  (2024-12-23T09:26:38Z)
• VideoLLaMB: Long Streaming Video Understanding with Recurrent Memory Bridges [39.666361965650836]
  VideoLLaMB is a framework for long video understanding.<n> SceneTiling algorithm segments videos into coherent semantic units.<n>VideoLLaMB processes up to 320 frames using a single Nvidia A100 GPU.
  arXiv  Detail & Related papers  (2024-09-02T08:52:58Z)
• Reuse and Diffuse: Iterative Denoising for Text-to-Video Generation [92.55296042611886]
  We propose a framework called "Reuse and Diffuse" dubbed $textitVidRD$ to produce more frames following the frames already generated by an LDM. We also propose a set of strategies for composing video-text data that involve diverse content from multiple existing datasets.
  arXiv  Detail & Related papers  (2023-09-07T08:12:58Z)
• Let's Think Frame by Frame with VIP: A Video Infilling and Prediction Dataset for Evaluating Video Chain-of-Thought [62.619076257298204]
  We motivate framing video reasoning as the sequential understanding of a small number of video reasonings. We introduce VIP, an inference-time challenge dataset designed to explore models' reasoning capabilities through video chain-of-thought. We benchmark GPT-4, GPT-3, and VICUNA on VIP, demonstrate the performance gap in complex video reasoning tasks, and encourage future work.
  arXiv  Detail & Related papers  (2023-05-23T10:26:42Z)
• No frame left behind: Full Video Action Recognition [26.37329995193377]
  We propose full video action recognition and consider all video frames. We first cluster all frame activations along the temporal dimension. We then temporally aggregate the frames in the clusters into a smaller number of representations.
  arXiv  Detail & Related papers  (2021-03-29T07:44:28Z)
• HERO: Hierarchical Encoder for Video+Language Omni-representation Pre-training [75.55823420847759]
  We present HERO, a novel framework for large-scale video+language omni-representation learning. HERO encodes multimodal inputs in a hierarchical structure, where local context of a video frame is captured by a Cross-modal Transformer. HERO is jointly trained on HowTo100M and large-scale TV datasets to gain deep understanding of complex social dynamics with multi-character interactions.
  arXiv  Detail & Related papers  (2020-05-01T03:49:26Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.