Emergent Temporal Correspondences from Video Diffusion Transformers

• URL: http://arxiv.org/abs/2506.17220v2
• Date: Mon, 23 Jun 2025 02:14:32 GMT
• Title: Emergent Temporal Correspondences from Video Diffusion Transformers
• Authors: Jisu Nam, Soowon Son, Dahyun Chung, Jiyoung Kim, Siyoon Jin, Junhwa Hur, Seungryong Kim,
• Abstract summary: We introduce DiffTrack, the first quantitative analysis framework designed to answer this question.<n>Our analysis reveals that query-key similarities in specific, but not all, layers play a critical role in temporal matching.<n>We extend our findings to motion-enhanced video generation with a novel guidance method that improves temporal consistency of generated videos without additional training.
• Score: 30.83001895223298
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Recent advancements in video diffusion models based on Diffusion Transformers (DiTs) have achieved remarkable success in generating temporally coherent videos. Yet, a fundamental question persists: how do these models internally establish and represent temporal correspondences across frames? We introduce DiffTrack, the first quantitative analysis framework designed to answer this question. DiffTrack constructs a dataset of prompt-generated video with pseudo ground-truth tracking annotations and proposes novel evaluation metrics to systematically analyze how each component within the full 3D attention mechanism of DiTs (e.g., representations, layers, and timesteps) contributes to establishing temporal correspondences. Our analysis reveals that query-key similarities in specific, but not all, layers play a critical role in temporal matching, and that this matching becomes increasingly prominent during the denoising process. We demonstrate practical applications of DiffTrack in zero-shot point tracking, where it achieves state-of-the-art performance compared to existing vision foundation and self-supervised video models. Further, we extend our findings to motion-enhanced video generation with a novel guidance method that improves temporal consistency of generated videos without additional training. We believe our work offers crucial insights into the inner workings of video DiTs and establishes a foundation for further research and applications leveraging their temporal understanding.

Related papers

• Tracktention: Leveraging Point Tracking to Attend Videos Faster and Better [61.381599921020175]
  Temporal consistency is critical in video prediction to ensure that outputs are coherent and free of artifacts.<n>Traditional methods, such as temporal attention and 3D convolution, may struggle with significant object motion.<n>We propose the Tracktention Layer, a novel architectural component that explicitly integrates motion information using point tracks.
  arXiv  Detail & Related papers  (2025-03-25T17:58:48Z)
• Understanding Long Videos via LLM-Powered Entity Relation Graphs [51.13422967711056]
  GraphVideoAgent is a framework that maps and monitors the evolving relationships between visual entities throughout the video sequence.<n>Our approach demonstrates remarkable effectiveness when tested against industry benchmarks.
  arXiv  Detail & Related papers  (2025-01-27T10:57:24Z)
• Enhancing Long Video Generation Consistency without Tuning [92.1714656167712]
  We address issues to enhance the consistency and coherence of videos generated with either single or multiple prompts.<n>We propose the Time-frequency based temporal Attention Reweighting Algorithm (TiARA), which judiciously edits the attention score matrix.<n>For videos generated by multiple prompts, we further uncover key factors such as the alignment of the prompts affecting prompt quality.<n>Inspired by our analyses, we propose PromptBlend, an advanced prompt pipeline that systematically aligns the prompts.
  arXiv  Detail & Related papers  (2024-12-23T03:56:27Z)
• DINTR: Tracking via Diffusion-based Interpolation [12.130669304428565]
  This work proposes a novel diffusion-based methodology to formulate the tracking task. Our INterpolation TrackeR (DINTR) presents a promising new paradigm and achieves a superior multiplicity on seven benchmarks across five indicator representations.
  arXiv  Detail & Related papers  (2024-10-14T00:41:58Z)
• TrackDiffusion: Tracklet-Conditioned Video Generation via Diffusion Models [75.20168902300166]
  We propose TrackDiffusion, a novel video generation framework affording fine-grained trajectory-conditioned motion control. A pivotal component of TrackDiffusion is the instance enhancer, which explicitly ensures inter-frame consistency of multiple objects. generated video sequences by our TrackDiffusion can be used as training data for visual perception models.
  arXiv  Detail & Related papers  (2023-12-01T15:24:38Z)
• Self-Supervised Video Representation Learning via Latent Time Navigation [12.721647696921865]
  Self-supervised video representation learning aims at maximizing similarity between different temporal segments of one video. We propose Latent Time Navigation (LTN) to capture fine-grained motions. Our experimental analysis suggests that learning video representations by LTN consistently improves performance of action classification.
  arXiv  Detail & Related papers  (2023-05-10T20:06:17Z)
• Transform-Equivariant Consistency Learning for Temporal Sentence Grounding [66.10949751429781]
  We introduce a novel Equivariant Consistency Regulation Learning framework to learn more discriminative representations for each video. Our motivation comes from that the temporal boundary of the query-guided activity should be consistently predicted. In particular, we devise a self-supervised consistency loss module to enhance the completeness and smoothness of the augmented video.
  arXiv  Detail & Related papers  (2023-05-06T19:29:28Z)
• Leaping Into Memories: Space-Time Deep Feature Synthesis [93.10032043225362]
  We propose LEAPS, an architecture-independent method for synthesizing videos from internal models. We quantitatively and qualitatively evaluate the applicability of LEAPS by inverting a range of architectures convolutional attention-based on Kinetics-400.
  arXiv  Detail & Related papers  (2023-03-17T12:55:22Z)
• Time Is MattEr: Temporal Self-supervision for Video Transformers [72.42240984211283]
  We design simple yet effective self-supervised tasks for video models to learn temporal dynamics better. Our method learns the temporal order of video frames as extra self-supervision and enforces the randomly shuffled frames to have low-confidence outputs. Under various video action recognition tasks, we demonstrate the effectiveness of our method and its compatibility with state-of-the-art Video Transformers.
  arXiv  Detail & Related papers  (2022-07-19T04:44:08Z)

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.