AIAny
Icon for item

Hunyuan3D-Buffalo 1.0: A Unified Multimodal Model for Scalable 3D Generation, Understanding, and Editing

Provides a unified multimodal framework for large-scale 3D understanding, text-to-3D generation, and instruction-guided 3D editing. Trains on an 87M-sample 3D multimodal corpus (25M understanding, 50M generation, 12M editing) and pairs a vision-language model with a diffusion-based 3D synthesizer to preserve structure and enable part-aware edits; suited for researchers building text-driven 3D asset pipelines but requires large compute and data.

Introduction

Most 3D modeling research is split between isolated generation, understanding, or editing tasks and is limited by the scarcity of large, geometrically consistent multimodal editing data. The core insight behind this work is that joint training on a purpose-built, large-scale 3D multimodal corpus plus an architecture that separates semantic reasoning (VLM) from high-fidelity synthesis (DiT) yields mutual gains across generation, understanding, and structure-preserving editing.

Key Findings
  • Scale and data engineering matter: the authors construct an 87M-sample 3D multimodal corpus (25M understanding samples, 50M text→3D pairs, 12M editing pairs) to close the data bottleneck for unified training. This enables capability transfer between tasks.
  • Hybrid architecture: a vision-language model supplies semantic, structural, and spatial conditions for a specialized Hunyuan3D DiT diffusion decoder, separating multimodal reasoning from synthesis so each component leverages pretrained priors effectively.
  • Editing and part generation: editing conditions the diffusion process on source-object representations to preserve unedited regions and overall structure; part-level generation is supported by text-grounded conditioning.
  • Empirical outcomes: reported results show state-of-the-art or leading performance on text-to-3D generation and 3D editing benchmarks, while also improving understanding and part-generation, indicating unified training benefits cross-task performance.
Who it's for and tradeoffs

Great fit if you are a researcher or engineering team building text-driven 3D asset pipelines, exploring unified multimodal 3D models, or needing controllable, structure-preserving 3D edits and part-aware synthesis.
Look elsewhere if you need an out-of-the-box lightweight solution: the approach relies on very large synthetic/augmented datasets and substantial compute for training, so replication or production deployment without access to similar data and resources will be costly.

Method snapshot

The practical takeaway is a two-part design: (1) a VLM for rich multimodal conditioning and spatial/semantic grounding, and (2) a diffusion-based 3D generator (Hunyuan3D DiT) that accepts those conditions plus optional source-object encodings for editing. The paper emphasizes dataset construction and conditioning strategies as central engineering contributions rather than novel single-module algorithms.

Information

  • Websitearxiv.org
  • OrganizationsTencent
  • AuthorsJunliang Ye, Kenkun Liu, Guocun Wang, Yang Li, Yansong Qu, Chunshi Wang, Jingwei Xu, Yunhan Yang, Zibo Zhao, Jiachen Xu
  • Published date2026/08/03

More Items

Introduces WorldExam, a diagnostic benchmark that evaluates controllable video world models across four levels from visual quality to inherent world reactivity. Covers 1,474 cases across eight tasks and supports camera-, action-, and language-driven paradigms, measuring scene-conditioned reactions beyond explicit instructions.

Estimates the visually attributable portion of a privileged teacher’s next-token corrections and reconstructs student-anchored training targets for multimodal on-policy distillation. Uses counterfactual teacher queries and a signed proxy to raise supported tokens and suppress refuted ones, improving fine-grained visual knowledge transfer across model scales.

Generates polygonal meshes from images using flow matching for fast, native mesh synthesis. Decodes vertices, edge connectivity, and face winding in one parallel pass, preserves artist-authored topology without vertex quantization or welding, supports a user-set vertex budget for face-count control, and completes image-to-mesh in ~6s median.