Vision & Video

Meshy T2 proposes a faster alternative to autoregressive mesh decoding by generating native polygon meshes with flow matching and per-vertex continuous latent tokens.

Pippa is pitching itself as a more ethical AI video platform by paying artists when users generate work in their styles. But its approach still depends partly on open models trained on broadly scraped online content.

VideoCoCo introduces a dual-engine framework that converts text prompts into executable Blender programs before transforming simulated drafts into photorealistic videos. The goal is to improve physical consistency in text-to-video generation.

Netflix’s ID-V2V targets identity-preserving video restylization: propagate scene, lighting, and style edits from keyframes across a source video without losing facial likeness, expressions, gaze, or lip sync.

O-VAD is a training-free framework for industrial video anomaly detection that tracks objects through time and reasons over their state trajectories. Its central claim is that frontier VLMs often fail in industrial settings because they are not given object-level evidence.

The paper tackles a practical weakness in long-horizon generative rendering: when a camera returns to a previously seen place, an autoregressive video model may redraw it differently. The proposed method restores historical latent chunks and uses 3D correspondences to guide attention, without post-training.

SANA-Video 2.0 proposes a hybrid video diffusion Transformer that keeps most of the efficiency benefits of linear attention while periodically restoring full softmax interactions. The paper reports competitive quality on a single H100, with notable speedups for longer and higher-resolution video generation.

Self Gradient Forcing addresses a missing training signal in autoregressive video diffusion: future losses rarely teach earlier generated latents how to become better KV memory. The proposed two-pass strategy restores that supervision without backpropagating through an entire long rollout.

TimeLens2 reframes video temporal grounding as an interval-set prediction problem, enabling multimodal models to identify not only what happens in a video but when the supporting evidence appears. Its 2B, 4B, and 8B variants show substantial gains over their Qwen3-VL backbones.

FlowMimic tackles a common bottleneck in video editing: expensive data collection and limited task diversity. Its key idea is to transform image editing samples into video editing samples in real time, while aligning image and video capabilities through mutual mimicry.

HOMIE targets human-object centric video personalization, aiming to preserve subject identity while improving interaction accuracy. It integrates multimodal model knowledge more effectively into video generation.

REBASE introduces a training-free framework for one-shot in-context segmentation that suppresses misleading background correspondences. By projecting features away from a reference-background subspace, it produces cleaner matches for promptable segmenters such as SAM.

VideoChat3 targets three persistent problems in open-source video MLLMs: weak generalization, high compute cost, and incomplete openness. With a 4B-parameter design, it aims to balance broad video understanding with practical efficiency.

A new arXiv paper presents BitMind Forensics, a deepfake detection system trained through an open adversarial competition that continually refreshes its data distribution. Across 19 public benchmarks, the dated BMF snapshot reports strong results on in-the-wild images, video, and AI-generated media.

VideoRAE explores whether frozen video foundation model representations can become compact, reconstructable, and generation-friendly video latents. The work targets a key limitation of conventional 3D-VAEs: latents optimized mainly for pixel reconstruction.

A new arXiv paper studies a multi-expert OCR pipeline for historical Manchu documents with sharply different visual styles. The system routes each page to a specialist model using a lightweight page-level image classifier.

A new arXiv paper on the ABAW11 challenge argues that valence-arousal estimation, expression recognition, and action unit detection should not necessarily share the same fusion strategy.

A new arXiv paper proposes VisionScreen, a vision model that extends the Screening mechanism from language modeling to image recognition. Instead of assigning relative softmax weights to all patches, it aims to explicitly exclude low-relevance visual tokens.

A new arXiv paper proposes modeling choreography as a sequence of interpretable atomic movements rather than only as continuous motion. The approach separates dance planning from motion completion to improve structure, rhythm alignment, and controllability.

A new arXiv paper introduces Peak-End-Net, a lightweight framework for video aesthetic assessment inspired by the psychological peak-end rule. Instead of treating all frames equally, it emphasizes salient moments and the ending of a video.

A new arXiv paper introduces TCA-Net, a low-light image enhancement network centered on more reliable fusion between intensity and chromaticity streams. Its key idea is to replace fixed Top-K sparse attention with thresholded cross-attention.