Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Emergent Temporal Correspondences from Video Diffusion Transformers
Authors: Jisu Nam, Soowon Son, Dahyun Chung, Jiyoung Kim, Siyoon Jin, Junhwa Hur, Seungryong Kim
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We introduce Diff Track, the first quantitative analysis framework designed to answer this question. Diff Track 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 Di Ts (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 Diff Track in zero-shot point tracking, where it achieves state-of-the-art performance compared to existing vision foundation and self-supervised video models. |
| Researcher Affiliation | Collaboration | 1KAIST 2Korea University 3Google Deep Mind |
| Pseudocode | No | The paper describes methods using text and mathematical equations, but it does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | We will make our code and evaluation dataset publicly available. |
| Open Datasets | Yes | Diff Track 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 Di Ts (e.g., representations, layers, and timesteps) contributes to establishing temporal correspondences. We evaluate zero-shot tracking on two real-video datasets with precisely annotated tracks: TAP-Vid-DAVIS [19] and TAP-Vid-Kinetics [19], following [46]. We will make our code and evaluation dataset publicly available. |
| Dataset Splits | Yes | To systematically analyze their intricate interplays during video generation, we construct a dataset of prompt-generated video using a video backbone under analysis and obtain pseudo ground-truth motion tracks from an off-the-shelf tracking method [44]. Each dataset includes 50 text prompts with corresponding 50 videos (e.g. 480 × 720 resolution, 49 frames, generated by Cog Video X-2B [84]). We predefine a set of starting points p1 ∈ R N ×2 (in latent resolution) in the first frame of each video, where N is the number of points. We evaluate zero-shot tracking on two real-video datasets with precisely annotated tracks: TAP-Vid-DAVIS [19] and TAP-Vid-Kinetics [19], following [46]. |
| Hardware Specification | Yes | All experiments were conducted on an A6000 GPU. |
| Software Dependencies | No | The paper mentions software tools like Co Tracker [44], SAM [48], and RAFT [76], but does not provide specific version numbers for these or other key software components used in their implementation. |
| Experiment Setup | Yes | For zero-shot tracking evaluation [5], we used the most significant layer and timestep, identified by matching accuracy in Sec. 3.4 and Sec. B: l = 17, t = 1 for Cog Video X-2B, l = 16, t = 1 for for Cog Video X-5B, and l = 16, t = 1 for Hunyuan Video. For CAG, we used the top-3 dominant layers l = 13, 17, 21 for Cog Video X-2B and l = 15, 17, 18 for Cog Video X-5B, as identified by harmonic mean in Sec. 3.4 and Fig. A.1. Following [2], we applied the guidance at all sampling timesteps. |