PaperlayerView exact revision on arXiv

Source-linked AI summary

WonderZoom: Multi-Scale 3D World Generation

Jin Cao, Hong-Xing Yu, Jiajun Wu

arXiv:2512.09164v1cs.CVcs.AIcs.GR

TL;DR

Existing 3D world-generation methods remain limited to single-scale synthesis, and multi-scale generation lacks a representation suited to progressive content creation. WonderZoom combines scale-adaptive Gaussian surfels with a progressive detail synthesizer; it reports improved visual quality and prompt alignment over existing 3D- and video-based methods.

  • Problem

    Existing 3D world-generation methods are limited to single-scale synthesis, and multi-scale generation lacks a scale-adaptive representation for progressively creating finer content.

    Source: §1 · p. 2

  • Method

    WonderZoom uses dynamically updatable scale-adaptive Gaussian surfels and a progressive synthesizer to generate finer 3D content conditioned on coarser scales and user-specified regions and viewpoints.

    Source: §1 · p. 2

  • Results

    WonderZoom reports significant improvements over existing 3D-based and video-based methods in visual quality and prompt alignment.

    Source: §5 · p. 8

  • Takeaways & Limitations

    Users can interactively zoom into generated scenes to create new details, supporting content creation and virtual-world exploration across multiple scales.

    Source: §1 · p. 2, §5 · p. 8

    WonderZoom can struggle when zooming deeply into pure texture regions because it relies on semantic cues to guide what to generate next.

    Source: §5 · p. 9

Abstract

from arXiv · show

We present WonderZoom, a novel approach to generating 3D scenes with contents across multiple spatial scales from a single image. Existing 3D world generation models remain limited to single-scale synthesis and cannot produce coherent scene contents at varying granularities. The fundamental challenge is the lack of a scale-aware 3D representation capable of generating and rendering content with largely different spatial sizes. WonderZoom addresses this through two key innovations: (1) scale-adaptive Gaussian surfels for generating and real-time rendering of multi-scale 3D scenes, and (2) a progressive detail synthesizer that iteratively generates finer-scale 3D contents. Our approach enables users to "zoom into" a 3D region and auto-regressively synthesize previously non-existent fine details from landscapes to microscopic features. Experiments demonstrate that WonderZoom significantly outperforms state-of-the-art video and 3D models in both quality and alignment, enabling multi-scale 3D world creation from a single image. We show video results and an interactive viewer of generated multi-scale 3D worlds in https://wonderzoom.github.io/

1. Introduction

Existing 3D world-generation methods are limited to single scales, while multi-scale generation requires a representation that can grow as finer content is synthesized. WonderZoom combines scale-adaptive Gaussian surfels with progressive detail synthesis to generate new content from a single image across scales.

1. Introduction

  • Single-scale methods cannot generate coherent content spanning panoramic vistas to microscopic details, limiting interactive exploration and content creation.The challenge is a scale-adaptive 3D representation for generation, where finer content is synthesized progressively rather than known upfront.Source: §1 · p. 2
  • WonderZoom combines dynamically updatable scale-adaptive Gaussian surfels with a progressive synthesizer that creates fine-grained 3D structures conditioned on coarser scales and user-specified regions and viewpoints.The representation can add arbitrary detail levels without re-optimization, while the synthesizer generates content coarse-to-fine.Source: §1 · p. 2
  • Users can zoom into any region to autoregressively synthesize previously nonexistent details, from landscapes to microscopic surface features.Unlike multi-resolution rendering that reveals pre-existing details, the system generates new content on demand.Source: §1 · p. 2
  • Across natural environments, villages, and urban scenes, WonderZoom reports consistent quality across scale transitions and significantly better perceptual quality and prompt alignment than state-of-the-art video and 3D models.The paper presents the approach as multi-scale 3D world generation from a single image.Source: §1 · p. 2

2. Related Work

Prior work supports single-scale 3D scene generation, multi-scale rendering or reconstruction with known content, and controllable synthesis in 2D or through refinement. WonderZoom is positioned as extending controllable, progressive multi-scale synthesis to 3D.

3D World Generation

  • Earlier 3D generation methods expanded from limited novel-view synthesis to connected scenes, rooms, and cities, but generate at a single spatial scale aligned with their input.WonderZoom instead lets users zoom into regions to synthesize new content at finer scales.Source: §2 · p. 2

Multi-scale 3D Representations

  • Traditional Level-of-Detail and neural hierarchical representations handle pre-authored or fully supervised multi-scale content, whereas generation must synthesize finer content progressively without pre-existing data.WonderZoom addresses this gap with a dynamically refined representation integrated into a progressive generation pipeline.Source: §2 · p. 3

Controllable Content Synthesis

  • Controllable video generation, superresolution, and hierarchical image generation do not provide the same 3D cross-scale synthesis: they lack scale-spanning training data, refine existing content, or remain limited to 2D.WonderZoom combines controllable generation with interactive zooming to synthesize new content across environmental and microscopic scales.Source: §2 · p. 3

3. Approach

WonderZoom builds a nested 3D scene by progressively synthesizing finer content from user prompts and zoom viewpoints. Scale-adaptive Gaussian surfels accommodate those additions, while the detail synthesizer creates and registers new-scale content and auxiliary views.

3. Approach

  • WonderZoom starts from an image-reconstructed scene, then generates nested, finer-scale scenes from user prompts and camera viewpoints that zoom into regions of interest.Each new scene contains finer-scale content within the preceding scene, rather than merely revealing pre-existing geometry.Source: §3 · p. 3

3.1. Scale-adaptive Gaussian Surfels

  • New surfels are appended at each scale without changing existing surfels, allowing the scene representation to grow without global re-optimization.Scale-aware opacity modulation controls overlapping surfels for real-time rendering, with smooth transitions between adjacent scales.Source: §3.1 · pp. 4–5

3.2. Progressive Detail Synthesizer

  • The progressive detail synthesizer renders a coarse zoomed view, adds semantic context from the previous scale, and applies super-resolution followed by prompt-guided editing.This combines enhancement of existing structures with insertion of new content specified by the user.Source: §3.2 · pp. 5–6
  • A depth estimator is aligned to rendered geometry at defined target-depth pixels, then refined segment-wise and for newly added structures to maintain geometric consistency.The method also uses Grounded SAM to isolate edited structures and estimate their depth relative to surrounding geometry.Source: §3.2 · p. 6
  • Neighboring-view images are synthesized from the partial scene and used with the new-scale image and depth to optimize a more complete 3D scene.These auxiliary views extend reconstruction beyond the single target viewpoint.Source: §3.2 · pp. 5–6

4. Experiments

The experiments compare WonderZoom with 3D and camera-controlled video baselines using qualitative and quantitative evaluations, then ablate three components. WonderZoom leads across alignment, novel-view quality, aesthetics, and human preference.

4. Experiments

  • Evaluation uses 32 generated scenes from 8 input images, four new-scale scenes per example, and fixed camera paths and prompts across methods.Source: §4 · p. 6

4.1. Comparison

  • State-of-the-art 3D scene-generation and controllable video methods cannot create multi-scale scenes: 3D methods produce blurry zoom-in views, while camera-controlled video can zoom but has less precise control than explicit 3D methods.Source: §4.1 · pp. 7–8
  • WonderZoom outperforms all baselines in alignment, novel-view quality, aesthetics, and human preference.Source: Table 1 · p. 8, Table 2 · p. 8, §4.1 · p. 8

4.2. Ablation study

  • Without scale-aware opacity modulation, rendering becomes blurry and computationally intractable for real-time multi-scale rendering; WonderZoom uses less GPU memory and renders faster.Source: §4.2 · p. 8, Figure 5 · p. 8, Table 3 · p. 8
  • Removing depth registration produces significant shape distortion in newly generated details from novel views; registration alleviates the artifact.Source: §4.2 · p. 8, Figure 6 · p. 8
  • Removing auxiliary view synthesis leaves regions missing, showing that it is critical for generating a complete 3D scene.Source: §4.2 · p. 8, Figure 7 · p. 8

5. Conclusion

WonderZoom generates multi-scale 3D worlds from a single image, letting users zoom into regions and synthesize new details while maintaining cross-scale consistency and real-time rendering. It improves visual quality and prompt alignment over existing 3D- and video-based methods, but can struggle when extreme zooming reaches texture regions without semantic cues.

5. Conclusion

  • WonderZoom generates multi-scale 3D worlds from a single image, synthesizing new details as users zoom into regions while maintaining cross-scale consistency and real-time rendering.The approach uses scale-adaptive Gaussian surfels and a progressive detail synthesizer.Source: §5 · p. 8
  • Experiments show improvements over existing 3D- and video-based methods in visual quality and prompt alignment.Source: §5 · p. 8
  • Extreme zooming into pure-texture regions can be difficult because the synthesizer relies on semantic cues to guide the next scale.The paper suggests texture-specific priors or procedural generation as future directions when semantic cues are insufficient.Source: §5 · p. 9

A. Algorithm

The appendix provides WonderZoom’s algorithm in Algorithm 1.

A. Algorithm

B. Additional Results

Additional visual results in Figures 9–12 are presented to demonstrate WonderZoom’s visual quality relative to baselines.

B. Additional Results

  • Figures 9, 10, 11, and 12 provide additional visual results intended to show WonderZoom outperforming baselines in visual quality.Source: B. Additional Results · p. 12

C. Failure Cases

Repeated zooming into a branch cluster can reduce the scene to texture without semantic cues, making further refinement unreliable and eventually preventing multi-scale world generation.

C. Failure Cases

  • Repeated zooming into a branch cluster can leave only texture patterns, with no recognizable branches or leaves.Source: C. Failure Cases · p. 12
  • Because WonderZoom uses current-scale image semantics to infer next-scale content, texture-only regions leave further refinement under-constrained and unreliable.Source: C. Failure Cases · p. 12
  • The failure arises when zoomed regions no longer contain semantic information; it does not occur while recognizable structure remains.Source: C. Failure Cases · p. 12

D. Additional Details

The additional details describe WonderZoom’s interactive rendering and progressive synthesis workflow, implementation settings, human preference study, and a failure case at excessive zoom depth.

D. Additional Details

  • Images are processed at 720 × 1088; GPT-4V extracts semantic context, and each new scale typically multiplies camera focal length by 8.INR-Harmonization is applied after image editing for shading consistency.Source: D. Additional Details · p. 12
  • The preference study recruited around 200 participants per comparison and used 2AFC responses across 24 pairs and 72 questions.The comparisons covered four baselines across six scenes; method order was randomized, and responses were anonymized.Source: D. Additional Details · p. 12
  • The control loop reconstructs an initial scene, renders continuously at a camera-dependent scale, and updates the camera from user controls.Progressive detail synthesis is triggered when a user zooms into a region and supplies an optional prompt.Source: D. Additional Details · p. 13
  • At a new scale, the system renders a coarse observation, extracts semantic context, applies super-resolution, and optionally edits the result using the user prompt.It then registers depth against the coarse scene and initializes surfels for the new scale.Source: D. Additional Details · p. 13
  • New-scale surfels are used to synthesize auxiliary views and are optimized with a loss combining L1 and D-SSIM terms.The algorithm specifies weights of 0.8 for L1 and 0.2 for D-SSIM.Source: D. Additional Details · p. 13
  • At excessive zoom depth into a tree, the view can collapse into texture-like patterns rather than meaningful branch structures.Source: Figure 13 · p. 18

Related papers

Loading 2512.09164v1…