Hunyuan3D-Buffalo, one model that builds, understands, and edits 3D from words
Hunyuan3D-Buffalo
One AI model now builds a 3D object from a sentence, and then changes it when you describe the change. Tencent’s Hunyuan team calls it Hunyuan3D-Buffalo. The same model also understands the shapes it makes, well enough to answer questions about them and point to their parts.
Four 3D jobs, now handled by one model
Until now, four 3D jobs each needed their own tool: understanding a 3D object and answering questions about it, building a new one from a prompt, editing one from an instruction, and pulling out a named part. Hunyuan3D-Buffalo does all four with one model. It comes from a research paper Tencent posted a few days ago, now trending on the AI paper boards. The weights and code are not out yet, so you can study how it works, not run it.
One shared brain instead of four separate tools
The real advance is coverage: one model covers all four jobs at once, where each used to need its own specialist system. It works because at the center sits one brain that reads pictures and words together, and Tencent taught it to read 3D too. They wired that brain to a second model that only builds 3D shapes, and trained it all on about 87 million 3D examples, a huge set, since good 3D data is the bottleneck.
One brain reads the shape and steers the generator
The brain starts from an existing model that already reads pictures and text. To give the brain a sense of 3D, a small encoder turns a shape into a short description: its geometry, the surface and which way it faces, plus its color. They also gave the brain extra words for writing coordinates, so it can point to a part by drawing a box around it. With just that, the brain can understand a shape and answer questions, all by generating text. To build something new, the brain’s read of your words passes through a small connector into a 3D generator, a model that starts from a rough, noisy blob and sharpens it into a shape. So one brain reads the shape and steers the building of one.
Editing is generation, holding everything else still
The last job, editing, reuses that same generator instead of a separate tool. Tencent reframes editing as generation: hand the generator the shape you already have, and tell it to build this again, changing only the part you asked about and keeping the rest. Building or pulling out a named part works the same way. So editing is just generation from an object you already have. One test shows why the design holds up. Tencent asked whether the part steering the edit really needs to understand the 3D shape, or could get by looking at a flat picture of it. Swapping the 3D-aware brain for a plain image reader raised the editing error by more than forty percent, and the result landed noticeably further from the target. Making the reading and the building one brain is what makes the editing good.
Picked most often in Tencent’s own four-way vote
On building 3D objects from text, Tencent ran a preference test: they showed people a prompt and the results from four systems and asked which was best. Buffalo’s result was picked about 57 percent of the time, against 18, 14, and 8 percent for the three earlier systems. That is a share of a four-way vote, in Tencent’s own study on a hundred prompts, not an absolute quality score and not an independent test. And the more 3D data they trained on, the higher that share climbed.
The paper is out; the weights are not
The paper is on arXiv, and the project page has the demo reel and the pipeline diagram. The links are on screen. The weights and code are not out yet, so this is a paper to read, not a model to download. If that changes, the license will be worth reading closely.
One brain builds it, reads it, and edits it
Four 3D jobs used to need four separate tools. Hunyuan3D-Buffalo does them with one model, because the same brain that reads a shape also drives the generator, and editing is just building it again while holding the rest still. The payoff is structural: building, understanding, and editing 3D become one problem instead of four.
