NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

From one photo to a 3D model in 0.5 seconds: VAST's Tripo and its US$50 million bet

VAST turned single-image 3D generation into a two-second job and closed a US$50 million round to push AI 3D into production pipelines.

2026-11-15 · 959 words · NeuroAI
From one photo to a 3D model in 0.5 seconds: VAST's Tripo and its US$50 million bet

A game artist drops a single screenshot of a wooden crate into a browser window and walks away. Two seconds later he returns to a clean, quadrilateral-meshed 3D object he can drop straight into his engine — no retopology, no UV headaches. For most of the industry's history, that crate would have cost an artist an afternoon. The two-second figure here refers to VAST's 2026 P1.0 production engine; the faster 0.5-second number in the title refers to the company's 2024 TripoSR research tool, a different stage of the same effort.

Tripo is the AI 3D (三维生成) product built by VAST, a Beijing company founded in 2023 and led by CEO Song Yachen (宋亚宸), CTO Liang Ding (梁鼎) and chief scientist Cao Yanpei (曹炎培), a Tsinghua-trained researcher who previously led generative 3D work at Tencent's ARC Lab. The company's bet is simple to state and hard to execute: make 3D asset creation as casual as typing a prompt.

The 0.5-second origin

VAST's public break came in March 2024, when it jointly open-sourced TripoSR with Stability AI. The model could turn one image into a 3D model in roughly 0.5 seconds, and it quickly became a go-to utility for 3D creators who wanted a fast first draft from a reference picture. That release was less a finished product than a statement of method: 3D generation could be fast, open and good enough to iterate on.

VAST then advanced the Tripo series — Tripo 2.0 in September 2024 and Tripo 2.5 in January 2025 — trained on tens of millions of high-quality native 3D assets, pushing generation speed, geometric accuracy and overall success rate with each iteration. In March 2025 the company launched a "Technology Open-Source Month," releasing eight projects across the technical chain, including TripoSG.

The architecture shift

TripoSG is the detail worth understanding. VAST describes it as the first Mixture-of-Experts (MoE) Transformer applied to 3D generation, built on a Rectified Flow foundation, with its 1.5-billion-parameter weights and inference code open on Hugging Face. The company says its output quality matches Tripo 2.0 while surpassing all existing open-source 3D generation projects, with strong generalization on complex composite objects.

The significance is architectural, not cosmetic. Most prior 3D models encoded shape as occupancy grids; TripoSG uses a signed-distance-function VAE that preserves finer geometry and generalizes across resolutions. Combined with a curated two-million-image-SDF training set, the approach trades raw parameter count for data quality and scaling discipline — the same lesson the language-model world learned earlier.

The 2-second production claim

The leap that reached production teams arrived in March 2026 with Tripo P1.0. Unlike earlier models that produced pretty-but-unusable "high-poly illusions," P1.0 generates a topologically clean, engine-ready low-poly mesh in about two seconds. VAST pairs it with a "Smart Mesh" workflow that accepts a prompt or a reference image and returns a model with sensible quad topology, so it can be animated or simulated inside engines like Unreal or tools like Maya without a human first cleaning up a tangle of triangles.

Cao Yanpei's framing is that AI should adapt to decades of industrial standards rather than force humans to repair what the model produced. The remaining hard problem, he acknowledges, is rigging and animation — giving assets bones and physical behavior — which is not yet fully production-ready but is advancing quickly. That honesty is unusual and useful: AI 3D (三维生成) is "usable" for props, environment kits and simple vehicles today, and still maturing for hero characters and facial animation.

The US$50 million vote

Weeks after P1.0, VAST closed a Series A of US$50 million (≈RMB 344 million), led by Alibaba and Hengxu Capital (恒旭资本, an SAIC-backed fund), with Yuanhe Puhua, BV Baidu Ventures and Orient Jiafu participating, and existing backers Primavera (春华创投) and the Beijing AI Industry Investment Fund adding capital. The raise signals that investors see 3D generation as infrastructure, not a demo — a large model (大模型) layer that sits beneath games, XR, industrial design and film.

VAST reports that its Tripo Studio has gathered about 6.5 million creators and generated nearly 100 million 3D models, with the majority of professional users outside China. Those figures are company-reported and unaudited, but the direction is consistent: a once-specialist skill is being turned into a utility.

Why this matters beyond gaming

The strategic prize is not prettier game assets. It is the lowering of a barrier that has kept 3D content scarce: most people on earth cannot author a 3D model, and that scarcity limits everything from e-commerce product views to training simulations to the open metaverse. If a photo or a sentence can yield a usable mesh in seconds, the population of "people who can make 3D" expands by orders of magnitude.

Honest limitations

This article relies on VAST's official blog, a Tencent News interview with chief scientist Cao Yanpei, and secondary venture-coverage for the funding round, model milestones and usage figures; I did not independently verify the 0.5-second or two-second timings on my own hardware. The "first MoE Transformer in 3D" and benchmark claims are the company's and were not reproduced by me. Creator and model counts are company-reported and unaudited. The piece does not assess content-labeling (内容标识) for generated assets, pricing for enterprise users, or how Tripo compares head-to-head with alternatives such as Tencent's Hunyuan3D or open tools on specific asset classes.

What readers can do now

  • If you build games, XR or product visualizations, trial Tripo's image-to-3D on your own reference photos and check whether the exported mesh needs manual cleanup before engine import.
  • For teams, pilot P1.0 on "filler" assets — crates, furniture, environment kits — where speed matters more than hero-grade detail.
  • Watch the open-source TripoSG release if you run your own infrastructure; an MoE 3D model you can self-host changes your per-asset cost calculus.

Related coverage

Cite this article

NeuroAI (脑机网), "From one photo to a 3D model in 0.5 seconds: VAST's Tripo and its US$50 million bet," 15 November 2026, https://neuroai.site/a/na-aigc-tripo-vast-3d-generation

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Generative Media” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories