Hunyuan3D-Buffalo 1.0: A Unified Multimodal Model for Scalable 3D Generation, Understanding, and Editing
Junliang Ye, Kenkun Liu, Guocun Wang, Yang Li, Yansong Qu, Chunshi Wang, Jingwei Xu, Yunhan Yang, Zibo Zhao, Jiachen Xu, Jiaao Yu, Lifu Wang, Zhihao Liang, Zhuo Chen, Chunchao Guo
cs.CV
2026-08-04
Tencent Hunyuan3D-Buffalo 1.0 unifies 3D understanding, text-to-3D, editing and part generation in one architecture; with an 87M-sample corpus it leads generation and editing benchmarks.
3D has long been split into disconnected pipelines: understanding (reading a 3D object), text-to-3D, editing existing objects, and part generation, each with its own model. The main obstacle to unification is data: large-scale, geometrically consistent, edit-supporting multimodal 3D data barely exists. Hunyuan3D-Buffalo 1.0 goes after both a single architecture that covers all four tasks and a self-built editing dataset to fill the gap.
Two components are joined:
The data engine is the centerpiece. It produces an 87M-sample corpus: 25M understanding samples, 50M text-to-3D pairs, and 12M editing pairs. The editing pairs come from a Nano3D-v2 pipeline with anchor-view selection, learned 3D-bbox localization, TRELLIS-voxel FlowEdit plus merging, LATTICE+NaTex refinement, and a final VLM filter.
On UniPart-Bench (part Q&A), SBERT 85.47 and SimCSE 89.06, the best. Text-to-3D human evaluation gives 55.2/57.1/56.6% preference across three dimensions, more than double Omni123 (17.5/21.0/18.4%). On Edit3D-Bench, average Chamfer Distance is 0.0091 versus Omni123 at 0.0684, an 86.7% reduction; average F1 is 0.6515 versus Steer3D at 0.2729, a 2.39x lead. A side finding: training on generation and understanding both feed back to improve editing, so unified training is mutually reinforcing.
A unified 3D model removes the cost of maintaining several specialized pipelines, and the mutual reinforcement among understanding, generation, and editing is the core selling point of this route. For teams in 3D assets, digital twins, and game or e-commerce 3D content, a single model that can see, create, and edit is closer to practical than a bag of single-purpose tools. The Nano3D-v2 pipeline for auto-generating editing data is itself reusable and eases the scarcity of 3D editing data.
The authors concede two points: single-stage high-quality geometry generation is unsolved and needs multi-stage DiTs (as in TRELLIS), and caption quality relies on Gemini and is noisy. A few more concerns: strong part-localization and editing results sit largely on the author-built UniPart-Bench and Edit3D-Bench, whose maturity and comparability await community checks; the VLM uses a Qwen-VL backbone, so geometric understanding is capped by that base; the 12M editing pairs are synthetic, and how well a model trained on them generalizes to messy real editing instructions is not fully validated.