Keep It Simple: Multi-Key Episodic Memory Retrieval for Ultra-Long Video Understanding

When videos extend from hours to days, directly processing them end-to-end becomes impractical for current Multi-modal Large Language Models (MLLMs). This ultra-long setting necessitates a two-stage paradigm: query-agnostic memory construction followed by retrieval-based inference. Prior work invests in complex memory construction to pre-model high-level relations in videos, despite not knowing the downstream query at build time. We instead prioritize high-recall retrievability during memory building, and defer query-specific, high-level relation composition to inference time. To this end, we propose MERIT(Multi-key Episodic Retrieval with Inference-time Temporal expansion), a simple yet effective agentic framework for ultra-long video understanding. First, we formulate an episodic multi-key representation that enables precise retrieval of fine-grained memories through a simple key-matching mechanism. Second, we introduce a neighbor filtering mechanism to capture broader semantic context without the massive computational overhead of global memory construction. This is achieved by expanding the temporal scope exclusively around the retrieved segments at inference time. By leveraging simple key-matching with this on-demand temporal expansion, MERIT achieves state-of-the-art performance across three long-video benchmarks: EgoLifeQA, LVBench, and Video-MME (Long).

Dual-Output Multi-Exposure HDR Reconstruction via SDR Fusion and Gain Map Inverse Tone Mapping

We propose DOME-HDR, a dual-output multi-exposure HDR reconstruction framework that jointly produces a perceptually balanced SDR image and a consistent HDR image via gain map inverse tone mapping. Given three bracketed LDR inputs, DOME-HDR first synthesizes a base SDR using a LoRA-adapted latent diffusion model. A dual cross-attention fusion module injects complementary structural and color cues from the under- and over-exposed images while anchoring on the mid exposure for stability. The synthesized SDR then guides HPGM, our HDR Prior-guided Gain Map network, to predict a spatially varying gain map for reliable dynamic-range expansion. We evaluate on Kalantari, Tel, and Challenge123 using both full-reference and no-reference metrics, where DOME-HDR achieves state-of-the-art HDR reconstruction quality; ablations further confirm the effectiveness of dual cross-attention and SDR-guided gain map estimation.

ATSplat: Compact Feed-forward 3D Gaussian Splatting with Adaptive Token Expansion

3D Gaussian Splatting (3DGS) achieves high-quality novel-view synthesis by optimizing freely placed primitives in 3D and adaptively densifying them in under-reconstructed regions. However, this scene-adaptive capacity allocation is largely lost in existing feed-forward 3DGS methods, which commonly regress Gaussians at input pixels and lift them along camera rays. Such pixel-aligned formulations make the number and placement of primitives depend on image resolution and input viewpoints rather than scene complexity, resulting in dense and often redundant Gaussian sets. We present ATSplat, a feed-forward 3DGS framework that restores the adaptive allocation capability of 3DGS optimization through Adaptive 3D Tokens. ATSplat first lifts coarse patch-level depth and camera cues into sparse 3D anchor tokens, forming a compact scaffold of the scene. Each token is then regressed into local Gaussians with learnable 3D offsets, decoupling primitive placement from input image grids. An Adaptive Token Expansion module predicts a token-level uncertainty score, supervised by rendering error maps, and selectively expands high-uncertainty tokens through learnable expansion layers. This sparse-to-adaptive formulation enables ATSplat to concentrate primitives in challenging regions while maintaining a compact representation. Experiments on two representative datasets, RealEstate10K and DL3DV, show that ATSplat achieves state-of-the-art rendering quality while reducing the number of Gaussians by more than $5.7\times$ compared with dense feed-forward 3DGS methods. From 12 input images at $512 \times 960$ resolution, ATSplat completes reconstruction in less than a second using a single commercial GPU, and renders high-quality novel views at 1136 FPS ($512 \times 960$) with only 311K Gaussians.

Coarse-to-fine Framework for Generative MEF via Implicit Neural Representation

Multi-exposure fusion (MEF) expands the luminance range beyond what a single exposure can capture. Combining images taken at different exposure levels requires handling geometric differences while naturally merging their complementary brightness information. It often demands generative completion where details are missing. Diffusion-based generative methods address these challenges, however, they are computationally expensive and struggle to preserve fine structures in saturated regions. We propose LIIFusion, a coarse-to-fine framework that balances fusion quality and efficiency in generative MEF. The coarse stage performs low resolution generative fusion, enhanced by an adaptive exposure correction that recovers structure lost in saturated over-exposed areas. The fine stage adapts a local implicit image function into a multi-exposure fusion function: conditioned on the HR OE/UE sources and the coarse output, it queries arbitrary target coordinates and fuses source evidence regardless of the HR input resolution. LIIFusion achieves up to 3.5$\times$ speed-up over existing generative methods while maintaining or improving structural fidelity and perceptual quality. We believe this framework provides an effective pathway toward making generative MEF more practical in real-world applications.

Scenes as Objects, Not Primitives: Instance-Structured 3D Tokenization from Unposed Views

A 3D scene is understood through its objects, not the primitives that compose them. Yet feed-forward reconstruction methods output dense, unstructured sets of points or Gaussians, leaving object-level structure to be recovered after the fact. We propose a feed-forward framework that decomposes a scene into instance-structured 3D token groups directly from unposed multi-view images -- compact object-centric units from which reconstruction, segmentation, and manipulation all follow. Each token group pairs an instance token capturing entity-level identity with anchor tokens that encode local geometry and appearance, which are decoded into a set of 3D Gaussians. This two-level factorization decouples object identity from local appearance, making object instances a native interface of the representation rather than a derived product. The token groups are learned through differentiable rendering with joint reconstruction and segmentation supervision, requiring no 3D annotations. Our feed-forward model surpasses per-scene optimization baselines in class-agnostic instance segmentation while remaining competitive in novel view synthesis. Beyond these metrics, the same token groups directly unlock instance-level scene editing -- removing, translating, or inserting objects by operating on their groups -- as well as efficient open-vocabulary 3D instance retrieval, where retrieval complexity scales with the number of instances rather than primitives.

Global Geometry Is Not Enough for Vision Representations

A common assumption in representation learning is that globally well-distributed embeddings support robust and generalizable representations. This focus has shaped both training objectives and evaluation protocols, implicitly treating global geometry as a proxy for representational competence. While global geometry effectively encodes which elements are present, it is often insensitive to how they are composed. We investigate this limitation by testing the ability of geometric metrics to predict compositional binding across 21 vision encoders. We find that standard geometry-based statistics exhibit near-zero correlation with compositional binding. In contrast, functional sensitivity, as measured by the input-output Jacobian, reliably tracks this capability. We further provide an analytic account showing that this disparity arises from objective design, as existing losses explicitly constrain embedding geometry but leave the local input-output mapping unconstrained. These results suggest that global embedding geometry captures only a partial view of representational competence and establish functional sensitivity as a critical complementary axis for modeling composite structure.

Seam360GS: Seamless 360° Gaussian Splatting from Real-World Omnidirectional Images

360-degree visual content is widely shared on platforms such as YouTube and plays a central role in virtual reality, robotics, and autonomous navigation. However, consumer-grade dual-fisheye systems consistently yield imperfect panoramas due to inherent lens separation and angular distortions. In this work, we introduce a novel calibration framework that incorporates a dual-fisheye camera model into the 3D Gaussian splatting pipeline. Our approach not only simulates the realistic visual artifacts produced by dual-fisheye cameras but also enables the synthesis of seamlessly rendered 360-degree images. By jointly optimizing 3D Gaussian parameters alongside calibration variables that emulate lens gaps and angular distortions, our framework transforms imperfect omnidirectional inputs into flawless novel view synthesis. Extensive evaluations on real-world datasets confirm that our method produces seamless renderings-even from imperfect images-and outperforms existing 360-degree rendering models.

Autoregressive Universal Video Segmentation Model

Recent video foundation models such as SAM2 excel at prompted video segmentation by treating masks as a general-purpose primitive. However, many real-world settings require unprompted segmentation that aims to detect and track all objects in a video without external cues, leaving today's landscape fragmented across task-specific models and pipelines. We recast streaming video segmentation as sequential mask prediction, analogous to language modeling, and introduce the Autoregressive Universal Segmentation Model (AUSM), a single architecture that unifies both prompted and unprompted video segmentation. Built on recent state-space models, AUSM maintains a fixed-size spatial state and scales to video streams of arbitrary length. Furthermore, all components of AUSM are designed for parallel training across frames, yielding substantial speedups over iterative training. On standard benchmarks (DAVIS17, YouTube-VOS 2018 & 2019, MOSE, YouTube-VIS 2019 & 2021, and OVIS) AUSM outperforms prior universal streaming video segmentation methods and achieves up to 2.5x faster training on 16-frame sequences.

Unsupervised Monocular 3D Keypoint Discovery from Multi-View Diffusion Priors

This paper introduces KeyDiff3D, a framework for unsupervised monocular 3D keypoints estimation that accurately predicts 3D keypoints from a single image. While previous methods rely on manual annotations or calibrated multi-view images, both of which are expensive to collect, our method enables monocular 3D keypoints estimation using only a collection of single-view images. To achieve this, we leverage powerful geometric priors embedded in a pretrained multi-view diffusion model. In our framework, this model generates multi-view images from a single image, serving as a supervision signal to provide 3D geometric cues to our model. We also use the diffusion model as a powerful 2D multi-view feature extractor and construct 3D feature volumes from its intermediate representations. This transforms implicit 3D priors learned by the diffusion model into explicit 3D features. Beyond accurate keypoints estimation, we further introduce a pipeline that enables manipulation of 3D objects generated by the diffusion model. Experimental results on diverse aspects and datasets, including Human3.6M, Stanford Dogs, and several in-the-wild and out-of-domain datasets, highlight the effectiveness of our method in terms of accuracy, generalization, and its ability to enable manipulation of 3D objects generated by the diffusion model from a single image.

Multi-Granular Spatio-Temporal Token Merging for Training-Free Acceleration of Video LLMs

Video large language models (LLMs) achieve strong video understanding by leveraging a large number of spatio-temporal tokens, but suffer from quadratic computational scaling with token count. To address this, we propose a training-free spatio-temporal token merging method, named STTM. Our key insight is to exploit local spatial and temporal redundancy in video data which has been overlooked in prior work. STTM first transforms each frame into multi-granular spatial tokens using a coarse-to-fine search over a quadtree structure, then performs directed pairwise merging across the temporal dimension. This decomposed merging approach outperforms existing token reduction methods across six video QA benchmarks. Notably, STTM achieves a 2$\times$ speed-up with only a 0.5% accuracy drop under a 50% token budget, and a 3$\times$ speed-up with just a 2% drop under a 30% budget. Moreover, STTM is query-agnostic, allowing KV cache reuse across different questions for the same video. The project page is available at https://www.jshyun.me/projects/sttm.