Category
Explore by categories
Argues a single web-scale generative video model handles vision tasks zero-shot the way LLMs handle language. Probes Veo 3 on segmentation, edge detection, image editing, physical and affordance reasoning, and puzzles like maze solving and symmetry.
Provides a diagnostic suite that audits video-understanding benchmarks to find samples solvable without visual or temporal input, filters those shortcuts, and produces a distilled video-native testbed that reveals major capability gaps in current Video-LLMs.
Performs training-free early-stage visual token compression inside the vision encoder to cut time-to-first-token (TTFT) and FLOPs for Video-LLMs. Introduces a decoupled spatial token selection strategy and reports up to 2.65× TTFT reduction and 61% FLOPs savings on LLaVA-OneVision-7B (NVIDIA A100) while preserving full-token accuracy — aimed at latency-sensitive video understanding.
Enables real-time streaming video-to-video editing (1280×704 @24 FPS) on a single RTX 5090 GPU. Uses a Hybrid Diffusion Transformer for balanced local/global modeling, Cycle‑Reverse Regularization for temporal consistency, and system-level mixed-precision and fused kernels to maximize throughput.
Decouples perception and reasoning for hours-long videos by streaming inputs into a three-tier Hierarchical Graph Memory and using an agentic Observation–Reason–Action retrieval loop; reduces reasoning context to ~2% of full video while improving benchmark accuracy.
Stores a persistent 3D scene cache directly in a diffusion model's latent space to produce temporally and spatially consistent videos. Constructs memory via depth-guided back-projection and queries it with direct latent-space warping — achieving large speed and memory gains versus pixel-space 3D baselines.
Continuously watches live video and autonomously decides each second whether to speak, stay silent, or delegate; released together with an 8B vision-first model, time-aligned interaction data, training recipe, and a deployable real-time system. Designed for vision-triggered, low-latency streaming scenarios and evaluated across six real-world streams.
Encodes and clones camera motion from reference videos to generate multi-shot videos — uses a visual "camera grid" to represent camera parameters, trains on million-scale grid–video pairs, and employs a hierarchical prompt-expansion agent to coordinate camera, subject, and action control for multimodal diffusion models.
Proposes chunk-level multimodal retrieval and chunk-adaptive reranking for retrieval-augmented generation on long egocentric videos; introduces V-RAGBench to decouple retrieval vs. generation evaluation and CARVE to run parallel retrievers and select per-chunk configurations.
Controllable long-horizon text/image-to-video generation that supports camera navigation, revisits, and promptable events across photorealistic and stylized domains. Introduces camera-aware positional encoding (E-PRoPE), memory-conditioned scene persistence, causal-forcing distillation, and RL alignment to retain camera control and reduce drift.
Serves interactive, long-lived streaming video-generation sessions by jointly scheduling session placement and GPU autoscaling to meet tight per-chunk latency. Combines migration-aware placement, load-driven autoscaling, coalesced chunk processing, GPU–CPU offloading and NCCL GPU–GPU migration; reports ~37% reductions in worst-case per-chunk latency and GPU operating cost.
Generates temporally grounded captions for dense multi-event videos by restructuring autoregressive token dependencies to enable lossless parallel decoding; introduces a latent global planning module and event-factorized parallel decoding to improve grounding accuracy and achieve large decoding speedups.