Tag
PPMem is a video world model that uses compact plane primitives as persistent 3D memory, jointly generating long coherent videos with explicit metric depth, normals, and exportable surfaces while outperforming prior 3D-grounded world models on DL3DV and RealEstate10K.
Gen4P 将预训练视频生成器改造为前馈式 4D 感知器，通过单次编码与按需交叉注意力解码，从单目视频中预测几何与运动，并在 3D 点跟踪、视频深度和相机位姿估计上达到 SOTA 或有竞争力的性能。
GeoFix提出了一种统一框架，通过单一几何奖励提升4D生成中的几何一致性和视觉质量，该奖励衡量静态区域的跨视角深度一致性，并针对视角重叠效应进行校准，实现了WorldScore 3D一致性提升39%和VBench质量提升4%。
Pumpire 提出了一个统一的基准测试，用于评估图像和视频级 3D 基础模型的点对点物理距离估计能力，并构建了包含 100 个真实场景、6400 帧的大型数据集 pumpire-6k，对 29 种基线配置进行了全面对比实验。
Iris-3B 是一个 3B 参数的像素空间文本到图像扩散模型，通过课程式预训练从零训练并与潜空间模型 FLUX.2 Klein 进行对比。结果显示像素空间生成先验在深度估计和图像修复任务上并无显著优势，但证明了像素空间预训练可扩展到与潜空间模型相当的画质。
Sperid Labs 发布 Iris-3B，一个 30 亿参数的像素空间生成模型，直接输出每个像素、无需 VAE 或潜空间，并可通过微调完成单目深度估计、图像修复与超分辨率等密集视觉任务。
Atlas is a tool that reconstructs spaces from a few photos and generates photorealistic RGB and depth data for robotic simulation, enabling robots to be trained and tested in more diverse environments.
This research investigates repurposing the backbone from YOLO26's depth estimation model for image deraining, demonstrating that depth-trained initialization provides a consistent improvement over random initialization in controlled experiments.
LingBot-Vision is a family of self-supervised vision transformer backbones for dense spatial perception, using masked boundary modeling to capture semantic and geometric structures for tasks like depth estimation and segmentation.
DreamX-Phi 1.0 is an action-conditioned video world model for robotic manipulation that predicts future observations using geometric attention encoding, depth estimation, and distillation, achieving top results in the WorldArena 2.0 Challenge.
ReViV is a unified framework for holistic egocentric 4D reconstruction that simultaneously reconstructs viewer (body, hand, gaze) and view (depth, camera trajectory) dynamics from a single monocular RGB video using a Masked Generative Egocentric Transformer, achieving state-of-the-art accuracy and efficiency.
Robbyant releases LingBot-Vision, a 1B-parameter vision model trained on boundaries that achieves better depth estimation than DINOv3-7B, with open weights.
This paper proposes masked depth modeling with sensor-validity masking, achieving best RMSE on 7 out of 8 masked/sparse depth benchmarks, with a controlled encoder-init study.
LingBot-Depth 2.0, trained on 150M samples, solves the longstanding problem of robots being blind to glass and transparent objects, achieving top performance on 12/16 depth benchmarks and halving depth error. Ant Group used it to significantly improve their robots' perception.
Robbyant, an embodied AI company under Ant Group, released LingBot-Vision, a self-supervised vision backbone family ranging from 21M to 1.1B parameters, under Apache-2.0. It matches or beats DINOv3 on several depth and segmentation benchmarks despite using less than one third of the training data, highlighting a push for open perception models.
Ant Group released LingBot-Vision, a family of DINO-style vision backbones in 4 sizes; the 0.3B ViT-L matches DINOv3-7B on NYUv2 depth with ~23x fewer parameters, showcasing significant efficiency gains.
This paper introduces masked boundary modeling, a self-supervised paradigm for vision pretraining that learns sub-pixel boundary representations to improve dense spatial perception. The resulting model, LingBot-Vision, demonstrates significant improvements in depth estimation and other downstream tasks, showing that boundary modeling is a scalable pretraining principle for spatially structured visual representations.
Introduces MultiDepth-3k, a benchmark to evaluate depth-layer preferences in monocular depth foundation models, and shows Laplacian Visual Prompting can alter reported depth layers, suggesting complementary geometric hypotheses exist across models.
MMDiff extends frozen diffusion transformers into multi-modal generative systems using lightweight decoders, achieving significant improvements in semantic segmentation and other perceptual tasks through multi-timestep feature fusion.
Bardienus Duisterhof introduces Modality Forcing, a recipe for post-training text-to-image (T2I) models that achieves state-of-the-art results on 4 out of 5 monocular depth estimation benchmarks.