video processing teams

Abstract: Video captioning models are increasingly evaluated with metrics designed for single-sentence image captions. We test whether those metrics recover human judgement on dense, multi-window video description, and find that they do not. Six frontier models captioned 196 short clips. Trained annotators post-edited every caption against a failure taxonomy fixed before annotation, once as a whole-clip overview and once window by window, and a quality-control reviewer checked each annotation, yielding 1,144 overview verdicts and 4,672 window verdicts. Reference-overlap metrics do not separate the captions reviewers accepted from those they rejected: BLEU-4 reaches an AUC of 0.510 and a CIDEr-style tf-idf cosine 0.528, and restricting references to reviewer corrections leaves BLEU-4 at 0.479 (95% CI 0.431-0.529), indistinguishable from chance. Review granularity changes what is detected. The same annotators record camera and shot errors on 2.7% of captions when reading the overview and on 12.8% when reading window by window, a 4.7x gap that holds across every robustness check we ran, on the mode most likely to need a near-total rewrite. Correct-overview rates span 54.1-72.7%, but only 4 of 15 pairwise gaps survive Bonferroni correction (5 after adjusting for reviewer and clip effects), and accuracy and derived severity order the models differently. Models with statistically indistinguishable accuracy have sharply different failure profiles, which no single scalar shows. We argue that caption evaluation should record what was wrong rather than only how often, and we report the taxonomy, the per-model profiles, the corpus properties that predict difficulty, and the code that regenerates every number in the paper.

Abstract: Feed-forward 3D foundation models have recently reshaped multi-view reconstruction, yet dynamic 4D understanding still often relies on large-scale motion supervision or hand-crafted, architecture-specific heuristics. We ask whether frozen 3D foundation model tokens already encode motion-relevant information that can be exposed by lightweight readout functions, rather than re-learning motion-specific features with high-capacity decoders. Through probing on three representative backbones, we find that static--dynamic separation is decodable from frozen token representations. Our rank-restricted and spectral analyses on $π^3$X reveal that the recovered motion cues are concentrated in a low-dimensional subspace. Building on this, we propose Probing-based Motion Distillation (PMD), which converts pairwise motion-state compatibility scores against a fixed anchor token into dense motion masks. Trained on only 160 YouTube-VOS videos using a single 16 GB GPU, PMD transfers zero-shot to DAVIS, SegTrackv2, and FBMS-59 without target-benchmark fine-tuning. With multi-view pretrained geometric backbones, PMD outperforms training-free heuristics and approaches substantially heavier supervised and self-supervised methods, suggesting that motion decoding from frozen 3D foundation models can be lightweight and data-efficient.

Abstract: Fixed-window Vision Transformer (ViT) decoders run at deployment resolution as overlapping tiles: every token in an overlap band is processed by several full decoder instances. In the 2.4B-parameter MiniMax-H3 video decoder, about half of all per-token computation is duplicated. Removing the tiles fails because attention, global tokens and coordinates are bound to them. TileShare keeps each released tile's attention and computes every other operation once per token. The released computation changes at one kind of step only: an overlap token's per-tile results are blended into one shared state with the released weights. Interventions show that this blend costs quality where attention reads the shared states, mainly in the first few layers, which TileShare therefore keeps per tile. Without training, keeping six layers per tile decodes 1.38-1.51x faster than batched and compiled tiling on two GPU types (up to 1.96x with FP8) and is equivalent to the released decoder in PSNR, LPIPS and DISTS on DAVIS videos. On unseen UVG videos with the setting frozen beforehand, it matches PSNR and DISTS and improves LPIPS. TileShare carries over to two further decoders and is again confirmed on unseen data. It is equivalent in PSNR and DISTS on MAGI-1. On the ViTok-v2 image autoencoder at 2K-4K, it matches overlapping tiling in PSNR, improves LPIPS and DISTS and decodes 1.5-1.8x faster. How many layers stay per tile is measured once per model (for decoders, without labels); we report it for 14 decoders and encoders. Optionally, two GPU-hours of self-distillation let MiniMax-H3 share every layer without a loss in PSNR, and a coordinate field trained while the network stays frozen brings fully shared CLIP-based open-vocabulary segmentation to parity with sliding windows on seven benchmarks.

Abstract: Training-free sparse attention is a practical route to efficient video diffusion transformer inference, yet existing methods typically fix the competition domain at each query row for every attention head. Prior work primarily addresses how to estimate token-level importance within this fixed domain; whether the domain itself should adapt to head heterogeneity remains under-specified. We introduce EG-Sparse, a training-free framework that routes low-entropy heads to a grouped cross-head pool while retaining per-row local selection for high-entropy heads, using only scores already materialized by the sparse pipeline. A controlled diagnostic fixes the ranking score and varies only the competition rule: grouped cross-head selection reduces low-entropy-head reconstruction error by 9.8%, whereas forcing high-entropy heads into an all-head global pool increases it by 7.9%. Across four configurations, EG-Sparse achieves higher video quality while reducing attention density and GPU denoising latency compared with strong training-free sparse-attention baselines. Mechanism ablations further show that selective domain adaptation outperforms uniform, random, and fixed domain-assignment policies. These results support competition-domain adaptation as a distinct design axis within centroid-guided training-free sparse attention.

Abstract: Vision-language models (VLMs) can answer questions about hour-long videos, but processing every frame is prohibitively expensive, even though the evidence for a question usually spans only a few seconds. Video agents, i.e., harness programs wrapped around a frozen VLM, address this by observing the video selectively, yet existing harnesses are hand-crafted by experts through slow build-and-test cycles. We propose VidHarness, a framework that automates harness design for cost-efficient long video understanding, in which a harness proposer iteratively evolves harnesses based on execution feedback from an evolution environment. To escape the local optima of greedy refinement, we organize the evolution as Monte Carlo tree search (MCTS), and to reduce the evaluation cost, we integrate uncertainty-aware multi-fidelity validation, which screens new harnesses on a few questions and promotes only the promising ones. Since the best harness varies with the frame budget, we further introduce a mixture-of-harness that routes each question to a harness specialized for its budget. VidHarness sets new state-of-the-art results on LongVideoBench, Video-MME, and Video-Holmes, outperforms the strongest hand-crafted video agent by up to $11.2$ points, and generalizes to the knowledge-intensive benchmarks Video-MMMU and MMVU with fewer than half of the frames of uniform sampling.

Abstract: Attention weights depend on token counts, which change with the representation of a signal. We study refinement symmetry: splitting a representation while preserving content, position, visible context, and total mass should preserve its contribution. Building on proportional and quadrature attention, we show that split invariance forces the local mass factor to be linear for any fixed positive attention kernel, provided that factor is nondecreasing. For changed representations, a physical coupling bounds attention error by separating feature change from weight reallocation. In Qwen2.5-Omni-7B, duplicating half the visual tokens threefold changes 255 of 3,586 MVBench answers under standard attention; measure weighting preserves every answer under matched visibility. Under natural frame resampling, it reduces distributional drift. At twofold merging of a frozen video encoding, a five-seed evaluation shows an all-partition-correct accuracy gain of 1.04 percentage points over global count weighting (average group mass) and 0.93 points over standard attention. The advantage over global count also holds on WorldSense but depends on the compression budget. The result is a representation principle with a measured benefit in robustness across partitions.

Abstract: Joint-embedding predictive architectures are unusually sensitive to how the input is masked: block masks work, scattered masks do not, and the explanations are empirical. We give a measurement account. A mask is a linear measurement, and in a compactly supported wavelet basis every atom whose support lies inside the hidden region falls in the measurement's null space and leaves no trace in the data. The JEPA loss asks only that the encoded context suffice for the target, so a target that a low-level prior can recover admits a shortcut, one the moving-average target encoder can make self-consistent. What removes the shortcut is the coarse-scale content the mask leaves unrecoverable, provided enough context stays within reach of each target. We score that content before training and test the account's distinctive predictions in 151 pre-training runs. On ImageNet-100, strip masks match blocks in area and contiguity yet are recoverable, and they land at 40.3% linear top-1, beside random masks at 40.8%, against 64.3% for blocks; within one geometry family, the placements that leave the least unrecoverable content lose 6.5 points to those that leave the most, over five seed pairs; pixel targets span 7 points where latent targets span 25; and against a frozen target the gap between random and block masks, 19 points on the same kind of GPU, closes to 1.5, so the geometry acts through the target the encoder produces for itself. On UCF101 the masking ratio decides which condition, content or reach, binds; removing whole frames, unrecoverable in space but recoverable from neighbouring frames, is worst at both ratios; and on V-JEPA's own masks, batching them intact instead of truncated changes little (36.0% against 35.1%), whereas making 100 target tokens inside the blocks visible lifts them to 48.7% and hiding 100 context tokens outside the blocks does not (33.7%).

Abstract: Training-free visual-token pruning often uses token importance, redundancy, or related selection criteria as proxies for safe removal. We show that these signals alone do not fully characterize removability, which is conditioned on both representation depth and the surrounding deletion set. Controlled interventions demonstrate that removing the same tokens at different depths produces substantially different downstream perturbations, while changing only the deletion context at a fixed depth alters candidate marginals and pruning-boundary decisions. These findings show that token importance alone cannot determine when a token is safely removable or how its removability changes under joint deletion. Motivated by this perspective, we propose CoRePrune, a training-free two-stage framework. Progressive Perturbation-Aware Visual Pruning refreshes deletion effects as visual representations evolve, while Set-Conditioned Refinement reevaluates candidate rescue benefits under the current deletion set after visual--text interaction. Across five multimodal large language model backbones covering standard images, high-resolution inputs, and video, CoRePrune preserves performance under aggressive token budgets. On Qwen3.5, with a final budget of 128 visual tokens, it retains 90.3% of dense-model performance while reducing aggregate prefill time by 51.0%.

Abstract: Over the past decade, a growing body of research has shown that $γ$-weak submodularity broadly arises in numerous subset selection tasks, including feature selection, neural network pruning, and video summarization. Despite its prevalence, maximizing a $γ$-weakly submodular function subject to a general matroid constraint remains challenging. To date, the only known approximation guarantee is the conservative $(1+1/γ)^{-2}$ factor established by \citet{chen2018weakly}. To improve upon this result, this paper proposes a novel algorithm called \MGPE, which repeatedly performs maximum-gain local exchanges through careful control of a non-homogeneous Poisson clock, and proves that this \MGPE\ can attain an approximation ratio arbitrarily close to $ρ_γ=1-\left(γ/(2-γ)\right)^{ \frac{γ^2}{2(1-γ)} }$. In sharp contrast to the previous guarantee, our obtained factor $ρ_γ$ not only strictly improves upon $(1+1/γ)^{-2}$ for every $γ\in(0,1]$, but also can asymptotically approach the optimal $(1-1/e)$-approximation for submodular maximization as $γ\to1$. Furthermore, we surprisingly find that when the matroid constraint reduces to a cardinality or the objective satisfies the stronger notion of $α$-weak DR-submodularity, \MGPE\ can automatically recover the tight approximation ratios of $1-e^{-γ}$ and $1-e^{-α}$, respectively. Here, $α\in(0,1]$ denotes the DR ratio.

Abstract: Long-context inference has become a defining capability of large language models, but exact dense attention remains costly due to its quadratic scaling with sequence length. Block-sparse attention offers a hardware-friendly alternative by routing each query block to a small set of relevant key blocks, yet accurate training-free block routing remains difficult. Existing routers often pool post-RoPE token representations, which entangles semantic aggregation with RoPE-induced geometry and attenuates local positional cues through high-frequency phase cancellation. To resolve this mismatch, we propose \textbf{Semantic-Geometric Decoupled Routing}, a training-free block routing framework that shifts semantic aggregation to the pre-RoPE space and reconstructs geometric bias with an offline structural prior and relative block distances. This decomposition yields an explicit closed-form block routing score without token-level search or post-hoc calibration. Experiments on long-context text and video tasks show that our method approaches full-attention accuracy across 4K--128K contexts, keeps routing overhead below 3.4 ms, and achieves a 5.03$\times$ speedup over FlashAttn at a 128K context length.

Abstract: Heavy rainfall severely degrades outdoor videos by corrupting high-frequency details and introducing motion blur, critically undermining the reliability of visual tasks. Recently, State Space Models (SSMs), particularly Mamba, have emerged as efficient alternatives for vision tasks with their linear complexity and ability to model long-range dependencies. However, when confronted with the poor visual representations in rainy videos, Mamba still faces difficulties in preserving the integrity of 2D spatial semantics and modeling 3D spatio-temporal correlations. To break these limitations, we introduce S3VD, a Semantic-Guidance Spatio-Temporal Scanning framework for video deraining, featuring two key innovations: Multi-Scale Semantic Fusion (MSSF) Module and Spatio-Temporal Scanning Fusion (STSF) Module. The former integrates temporal semantic priors from DINOv2 to guide precise feature representation and counteract the loss of local semantic context inherent to Mamba's 1D flatten operation, enhancing robustness against extreme degradation. The latter introduces a spatio-temporal scanning mechanism and devises a Decoupled-Gating Mamba (DG-Mamba) layer, which employs two independent gates to adaptively control preceding and subsequent contextual information within the input clip, optimizing intra-frame and inter-frame correlation modeling. Experiments on video deraining benchmarks demonstrate the superiority of S3VD, achieving state-of-the-art performance with an average 0.84 dB PSNR improvement over Mamba-based baselines.

Abstract: Preparing high quality training data for foundation models requires scalable pipelines that transform heterogeneous documents and videos into structured records. Such pipelines expand each parent item into an ordered and input dependent sequence of children, whose counts may be long tailed. GPUs should batch children across parents while preserving parent relationships, child order, completion status, and result routing. Existing systems either hide parallelism behind coarse grained jobs or expose flat records that force applications to manage lineage and regrouping. We present RayOrch, a programming model and distributed execution engine that preserves parent child relations throughout execution. Programs declare ordered variable cardinality expansions and matching gathers. The compiler validates each pair, while the runtime records child membership, immediate parents, immutable ordinals, and terminal states. Per Call FIFO Ready Queues batch ready children across parents. Gathers reconstruct results from declared membership and ordinals rather than batch boundaries or completion order. Parents can advance as soon as all required children become terminal. Typed parent scoped failures suppress undispatched siblings of the failed parent while allowing unrelated parents to continue. On NVIDIA H20 GPUs, RayOrch achieves 15.14 times speedup when scaling MinerU from 4 to 64 GPUs and 7.82 times speedup when scaling a video pipeline from 8 to 64 GPUs. It reduces end to end time by 13.1 percent versus Ray Data and 29.0 percent versus Daft on MinerU, and by 16.0 percent versus Ray Data on Docling. Code available at https://github.com/OpenDCAI/RayOrch .

Abstract: Generative modeling of time-dependent data is typically formulated on a discrete temporal grid, restricting supervision to the observed timestamps in the training data. We instead frame generation as continuous-time evolution on a learned data manifold. To this end, we leverage pretrained score-based models as geometric priors and learn a vector field that evolves data along score-induced interpolation paths. Because these dynamics follow transitions that respect the geometry learned by the score model, they support generation at arbitrary timestamps and temporal super-resolution beyond the discretization of the training data. Moreover, this geometric formulation allows us to train the vector field simulation-free through a regression objective. To improve long-horizon rollout robustness, we introduce an objective that promotes path-relative transverse exponential stability. While motivated by stability theory, it admits a practical interpretation as denoising score matching transverse to the interpolation path. Further, we extend the framework to a probabilistic setting that models a distribution over plausible future trajectories. We demonstrate the method on natural video and scientific dynamical data, including temporal super-resolution, PDE-based spatiotemporal fields, and molecular dynamics. Our results show that score-based priors provide a strong foundation for learning stochastic continuous-time generative dynamics.

Abstract: In sign language recognition, the isolated (ISLR) classification loss treats a single label as ground truth, as does the frame-level auxiliary classifier over pseudo-labels we add to continuous (CSLR) methods, which lack one. Stylistic variation blurs ISLR annotation and the lack of temporal boundaries in CSLR forces pseudo-labels; both are noisy. We therefore apply symmetric and generalized cross entropy (SCE, GCE), robust alternatives to cross entropy (CE) from image classification, not to connectionist temporal classification but to the preceding single-label classifier. On ASL Citizen with injected symmetric noise on three backbones (three seeds for ST-GCN), robust losses cost at most 2.5 pt when labels are clean and beat CE by 2.9-10.0 pt in all six conditions at noise rate 0.2, one of which only after q was re-selected on dev. GCE gains more, but its optimal q does not transfer across backbones, whereas one SCE setting works in all nine conditions; both vary 2-11 times more than CE across runs, so a favorable point estimate does not establish stability. For CSLR (PHOENIX-2014) we report no gain; our frame-level targets carry a systematic assignment bias, making that study a diagnosis of a single configuration. At lambda_aux = 25 the pseudo-label CE auxiliary raises word error rate above the no-auxiliary baseline on VAC, CorrNet and SlowFastSign, and GCE/SCE improve on CE by 1.7-3.2 pt (three of six conditions return below that baseline). However, the three losses differ by more than an order of magnitude in effective gradient at a common lambda_aux: matching the initial gradient shrinks the gap to 0.4-0.9 pt, and lowering the CE weight alone already beats that baseline, so neither the degradation nor the improvement can be separated from the effect of the weight. We use only symmetric noise; multi-seed evaluation covers only ST-GCN and VAC isolated.