You See it, You Got it: Learning 3D Creation on Pose-Free Videos at Scale

TL;DR

The paper presents See3D, a scalable approach to open-world 3D creation trained on pose-free, web-scale video data. It introduces WebVi3D, a large dataset curated from Internet videos, and a 2D-inductive visual-condition that enables multi-view diffusion without camera poses. See3D is integrated into a warping-based 3D generation framework with depth estimation, depth-scale alignment, and 3D Gaussian Splatting reconstruction to produce consistent novel views and 3D meshes. The approach demonstrates state-of-the-art performance on single and sparse-view reconstruction benchmarks and enables long-sequence view generation and editing in open-world scenarios. This work highlights the potential of leveraging massive unposed video data to learn robust 3D priors at scale, reducing dependence on costly 3D datasets.

Abstract

Recent 3D generation models typically rely on limited-scale 3D `gold-labels' or 2D diffusion priors for 3D content creation. However, their performance is upper-bounded by constrained 3D priors due to the lack of scalable learning paradigms. In this work, we present See3D, a visual-conditional multi-view diffusion model trained on large-scale Internet videos for open-world 3D creation. The model aims to Get 3D knowledge by solely Seeing the visual contents from the vast and rapidly growing video data -- You See it, You Got it. To achieve this, we first scale up the training data using a proposed data curation pipeline that automatically filters out multi-view inconsistencies and insufficient observations from source videos. This results in a high-quality, richly diverse, large-scale dataset of multi-view images, termed WebVi3D, containing 320M frames from 16M video clips. Nevertheless, learning generic 3D priors from videos without explicit 3D geometry or camera pose annotations is nontrivial, and annotating poses for web-scale videos is prohibitively expensive. To eliminate the need for pose conditions, we introduce an innovative visual-condition - a purely 2D-inductive visual signal generated by adding time-dependent noise to the masked video data. Finally, we introduce a novel visual-conditional 3D generation framework by integrating See3D into a warping-based pipeline for high-fidelity 3D generation. Our numerical and visual comparisons on single and sparse reconstruction benchmarks show that See3D, trained on cost-effective and scalable video data, achieves notable zero-shot and open-world generation capabilities, markedly outperforming models trained on costly and constrained 3D datasets. Please refer to our project page at: https://vision.baai.ac.cn/see3d

Figures (6)

• Figure 1: Benefiting from the proposed web-scale dataset WebVi3D, See3D enables both object- and scene-level 3D creation, including (text-) image-to-3D, sparse-view-to-3D, and 3D editing. It can also be used for Gaussian Splatting to extract meshes or render images.
• Figure 2: Overview of See3D. (a) We propose a four-step data curation pipeline to select multi-view images from Internet videos, forming the WebVi3D dataset, which includes $\sim$16M video clips across diverse categories and concepts. (b) Given multiple views, we corrupt the original data into corrupted images $c_{t}^{i}$ at timestep $t$ by applying random masks and time-dependent noise. We then reweight the guidance of $c_{t}^{i}$ and the noisy latent $x_{t}^{i}$ for the diffusion model to form visual-condition$v_{t}^{i}$ through a time-dependent mixture. (c) MVD model is capable of training at scale to generate multi-view images conditioned on $v_{t}^{i}$, without requiring pose annotations. Since $v_{t}^{i}$ is a task-agnostic visual signal formed through time-dependent noise and mixture, it enables the trained model to robustly adapt to various downstream tasks.
• Figure 3: (a-Row1): Dynamic content alters scene geometry across views; (a-Row2): Limited camera movement provides insufficient multi-view observations; (b) Our WebVi3D comprises with static scenes and diverse camera trajectories.
• Figure 4: See3D for Multi-View Generation: From iteratively generated views (brown camera), we randomly select a few anchor views (yellow stars) to guide the generation of target views along the gray camera trajectory. Keypoint matching is first performed to establish correspondences between the anchor views. Next, monocular depth estimation is applied to the latest anchor view, followed by our Iterative Sparse Pixel-Wise Depth Alignment to refine the depth and recover a dense map. This dense depth is then used to warp images along the gray camera viewpoints. Subsequently, the warped images and anchor images are combined and processed according to Eq.\ref{['eq:ct']} and Eq.\ref{['eq:mix']}, without random masking, forming the visual-condition, which guides MVD model to produce 3D-consistent target views. Finally, the gray camera turns to brown, guiding multi-view generation in the next iteration.
• Figure 5: Qualitative Comparison of Single/Sparse View Generation. Top three rows are results with a single view input. Bottom two rows are novel view renderings from 3DGS, where Ours is trained on dense multi-view generation given 3 views as input. Our method outperformed other baselines in capturing high-frequency details, such as text and stairs. Best viewed in zoom.