Muses: Designing, Composing, Generating Nonexistent Fantasy 3D Creatures without Training

TL;DR

Muses tackles the challenge of generating novel fantasy 3D creatures without training data by anchoring design in a 3D skeleton and executing a design–compose–generate pipeline. It introduces a graph-guided skeleton design, a SLAT-based content fusion that maps skeletal structure to a region-aware latent space, and a style-aware texture refinement that yields harmonized textures. The approach achieves state-of-the-art fidelity and text alignment while enabling skeleton-aware 3D editing, and it operates in a fully feed-forward, training-free manner. Together, these components enable flexible creation and editing of complex, out-of-domain 3D assets with broad applicability in gaming, VR, and animation.

Abstract

We present Muses, the first training-free method for fantastic 3D creature generation in a feed-forward paradigm. Previous methods, which rely on part-aware optimization, manual assembly, or 2D image generation, often produce unrealistic or incoherent 3D assets due to the challenges of intricate part-level manipulation and limited out-of-domain generation. In contrast, Muses leverages the 3D skeleton, a fundamental representation of biological forms, to explicitly and rationally compose diverse elements. This skeletal foundation formalizes 3D content creation as a structure-aware pipeline of design, composition, and generation. Muses begins by constructing a creatively composed 3D skeleton with coherent layout and scale through graph-constrained reasoning. This skeleton then guides a voxel-based assembly process within a structured latent space, integrating regions from different objects. Finally, image-guided appearance modeling under skeletal conditions is applied to generate a style-consistent and harmonious texture for the assembled shape. Extensive experiments establish Muses' state-of-the-art performance in terms of visual fidelity and alignment with textual descriptions, and potential on flexible 3D object editing. Project page: https://luhexiao.github.io/Muses.github.io/.

Figures (8)

• Figure 1: Generated nonexistent fantastic 3D creatures including animals, humanoids, and fictional characters by Muses. Driven by a 3D skeleton, Muses is able to design basic 3D structures, compose different concepts, and generate high-fidelity creative 3D assets. Although the content comes from different sources, the generated creatures contain harmonious geometry and textures across different styles.
• Figure 2: Compared with (a) methods that distill part-level affinity from 2D generative priors li2025dreambeast, (b) methods that lift creative 2D images wu2025less to 3D xiang2025structured, and (c) methods that perform part-level generation yang2025omnipart (where (c) is obtained by manually assembling the generated parts), Muses generates creatures that better preserve creative intent and achieve higher structural coherence.
• Figure 3: Overview of Muses. Our framework automates fantastic creature generation through a 3D skeleton-driven pipeline of design, composition, and generation. Given a text prompt, Stage I parses it into concepts, generates corresponding 3D assets $\{\mathbf{X}\}_{m=1}^{M}$ and skeletons $\{\mathbf{G}=(\mathbf{V},\mathbf{E})\}_{m=1}^{M}$, and uses graph classification with LLM-guided reasoning to produce a text-aligned skeleton $\dot{\mathbf{G}}$. In Stage II, this skeleton guides part assembly in a structured latent space (SLAT), yielding a composed latent code $\mathbf{Z}'$. In Stage III, $\mathbf{Z}'$ is decoded into a coarse 3D creature $\mathbf{X}'$, which guides geometry-invariant texture editing and undergoes a final style-conscious refinement to produce the detailed, harmonious output $\mathbf{X}"$. The entire pipeline is automatic, training-free, and feed-forward.
• Figure 4: Skeleton-guided concept design.
• Figure 5: Comparison with the state-of-the-art methods. Note that DreamBeast li2025dreambeast cannot handle contents with more than three animals, and OmniPart yang2025omnipart requires manual stitching. Our method generates fantastic 3D creatures with superior quality and textual alignment.