WildRelight is the first in-the-wild dataset designed to evaluate single-image relighting models, featuring high-resolution outdoor scenes captured under strictly aligned, temporally varying natural illuminations paired with high-dynamic-range environment maps. The benchmark demonstrates that state-of-the-art models trained on synthetic data suffer severe domain shifts when applied to real-world imagery. Leveraging the dataset’s temporal structure, the authors introduce a physics-guided inference framework combining Diffusion Posterior Sampling with Temporal Sampling-Aware Test-Time Adaptation, enabling synthetic models to self-supervise and align with real-world statistics on-the-fly.
arXiv:2502. 07531v5 Announce Type: replace-cross Abstract: Controllable image-to-video (I2V) generation transforms a reference image into a coherent video guided by user-specified control signals.
RelightFormer is a feed‑forward generative Transformer that performs single‑ and multi‑view image relighting without explicit intrinsic property estimation. It incorporates a latent illumination module that injects target environment maps into spatial features via cross‑attention, and uses permutation‑invariant positional encodings to process unordered multi‑view inputs symmetrically. Trained on the large Laval Objaverse Dataset, the model achieves state‑of‑the‑art visual and photorealistic relighting quality, and demonstrates strong zero‑shot generalization across various relighting tasks.
SpatialCrafter introduces a two‑stage framework for single‑image world modeling that first generates a global 3D proxy using a Point‑anchored Sparse Structure Flow module, then refines appearance with a Generative Deferred Refiner built on a video diffusion model. The method incorporates Parallel Geometry Injection and Proxy‑Aware Corruption training to integrate the proxy without disrupting the pretrained generative manifold, and it is evaluated on a newly constructed dataset of 115K scenes. Experiments demonstrate that SpatialCrafter outperforms existing approaches, reducing long‑term drift and maintaining consistency under rapid camera motion and extreme viewpoints.
Dynamic-Robust Photometric-Semantic Reconstruction for Open-Vocabulary 3D Scene Understanding introduces SPAR, a joint semantic‑geometric encoding architecture that isolates transient dynamic noise before latent space aggregation. The method couples motion estimation with multi‑view visual and semantic learning in a dynamic‑region‑aware end‑to‑end training paradigm, enabling the network to resolve motion conflicts and produce temporally stable scene representations. Experiments on the D‑RE10K benchmark show state‑of‑the‑art performance, achieving high PSNR values for novel view synthesis and an 88.5% mIoU for motion mask prediction in a self‑supervised setting.
arXiv:2608. 20212v1 Announce Type: new Abstract: High-fidelity removal of eyeglasses from video is a major challenge in facial attribute editing, as the underlying facial geometry is often obscured by complex refractive distortions and view-dependent specular reflections.
arXiv:2607. 02269v1 Announce Type: cross Abstract: Vision-Language Models (VLMs) have demonstrated immense promise in Spatio-Temporal Video Grounding (STVG).
The paper introduces GeoNeXt, a framework that repurposes pretrained video generative models for geometry estimation by framing it as a next‑frame prediction task. Unlike prior methods that either train separate depth/normal models or fine‑tune image diffusion backbones, GeoNeXt jointly models images and geometric targets, leveraging the structured knowledge of video models for more data‑efficient learning. Experiments show zero‑shot monocular depth and surface normal estimation that outperforms existing generative approaches and rivals discriminative state‑of‑the‑art methods while using far less training data.
The paper introduces Driving with DINO (DwD), a framework that uses Vision Foundation Module (VFM) features to bridge simulation and real-world domains for autonomous driving video generation. It addresses the consistency‑realism dilemma by projecting VFM features onto a principal subspace, dropping high‑frequency texture elements, and applying a Random Channel Tail Drop to preserve structural detail. Additional components— a learnable Spatial Alignment Module and a Causal Temporal Aggregator— enhance control precision, spatial alignment, and temporal stability, reducing motion blur and ensuring realistic, consistent outputs.
The paper introduces a benchmark called Shedding Light to evaluate how well generative image models understand and reproduce lighting. The benchmark tests models by asking them to inpaint a simple object, called a light probe, into real photographs and then compares the generated probe to the ground truth to assess lighting direction, colour, and radiance. The authors provide a scalable protocol and open-source code and data for systematic assessment of photometric accuracy in future models.
Recent advances in diffusion models have shown impressive performance in controllable image generation and dense prediction tasks. However, existing approaches typically treat diffusion-based controllable generation and dense prediction as separate tasks, overlooking the potential benefits of jointly modeling the heterogeneous distributions.
arXiv:2606. 16278v1 Announce Type: cross Abstract: Long-tail hazardous scenarios are essential for safety-oriented autonomous driving, yet they are difficult to collect and reproduce at scale.
SolarWM is an open foundation for building interactive video world models, offering a reconfigurable multi‑source data engine that unifies 1.43 million clips from 10 datasets into a consistent, frame‑aligned format. It provides a backbone‑native adaptation framework that preserves native representations of models ranging from 5 B to 33 B parameters, and a three‑stage training recipe combining bidirectional adaptation, teacher‑forced autoregressive initialization, and distribution‑matching distillation. The resulting causal models can interact in real‑time over rollouts from minutes to hours, trained only on 5‑second sequences, and the project releases data, pipeline, recipes, weights, and framework for reproducible research.