Image description DreamBeast: Distilling 3D Fantastical Animals with Part-Aware Knowledge Transfer

Runjia Li1, Junlin Han1, Luke Melas-Kyriazi1, Chunyi Sun2, Zhaochong An3, Zhongrui Gui1, Shuyang Sun1, Philip Torr1, Tomas Jakab1
1University of Oxford, 2Australian National University, 3University of Copenhagen

In 3DV 2025

Paper arXiv Code
mainimg

DreamBeast generates fantastical 3D animal assets composed of distinct parts.

TL;DR

  • We introduce DreamBeast, a new method for generating part-aware 3D asset efficiently.
  • Utilize a novel part-aware knowledge transfer mechanism.
  • Efficiently extract part-level knowledge from Stable Diffusion 3 into a Part-Affinity NeRF for instant generation from various camera views.
  • Render Part-Affinity maps from the Part-Affinity NeRF and modulate a multi-view diffusion model during score distillation sampling (SDS).
  • Improves the part-awareness and quality of generated 3D creatures with efficient computational costs.

3D Fantastical Animal Generation

Abstract

We present DreamBeast, a novel method based on score distillation sampling (SDS) for generating fantastical 3D animal assets composed of distinct parts.

Existing SDS methods often struggle with this generation task due to a limited understanding of part-level semantics in text-to-image diffusion models. While recent diffusion models, such as Stable Diffusion 3, demonstrate a better part-level understanding, they are prohibitively slow and exhibit other common problems associated with single-view diffusion models. DreamBeast overcomes this limitation through a novel part-aware knowledge transfer mechanism. For each generated asset, we efficiently extract part-level knowledge from the Stable Diffusion 3 model into a 3D part-affinity implicit representation. This enables us to instantly generate part-affinity maps from arbitrary camera views, which we then use to modulate the guidance of a multi-view diffusion model during SDS to generate 3D assets of fantastical animals.

DreamBeast significantly enhances the quality of generated 3D creatures with user-specified part compositions while reducing computational overhead, as demonstrated by extensive quantitative and qualitative evaluations.

Pipeline

mainimg

    1. Partially optimize a NeRF using standard SDS.
    2. Render multiple views of the partially optimized NeRF and input them into SD3 with a text prompt to create Part-Affinity maps via cross-attention.
    3. Train a Part-Affinity NeRF using these extracted maps.
    4. Freeze the optimized Part-Affinity NeRF; render both 3D asset NeRF and Part-Affinity NeRF from the same camera pose. And use the rendered Part-Affinity map to modulate cross and self-attention in MVDream, generating a part-aware 3D animal.

Part-Affinity Map

Results

mainimg

Comparison with baseline methods.

mainimg

Part-Affinity maps visualization.

Non-animal Part-aware Generation

mainimg mainimg mainimg

BibTeX

@misc{li2024dreambeastdistilling3dfantastical,
      title={DreamBeast: Distilling 3D Fantastical Animals with Part-Aware Knowledge Transfer}, 
      author={Runjia Li and Junlin Han and Luke Melas-Kyriazi and Chunyi Sun and Zhaochong An and Zhongrui Gui and Shuyang Sun and Philip Torr and Tomas Jakab},
      year={2024},
      eprint={2409.08271},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2409.08271}, 
}