VAST V・STAR: 3D Generation & World Models as the Next AI Research Frontier

The Rise of LLM-Driven 3D Creation

Since September 2024, top-tier large models — GPT-6 Astra, Claude Opus 5.5, DeepSeek V4.1-Flash-1017, Kimi K3.1 Pro — have demonstrated 3D creation as a new capability showcase: building rooms from a sentence, constructing mechanical devices with code, generating interactive 3D web pages. As coding and tool-use abilities improve, workflows that once required multiple professional software packages are being condensed into natural-language-driven pipelines.

When demos become real creation, concrete problems emerge: Can characters match reference images? Can mechanical parts move independently? Do generated assets integrate cleanly into game engines? Can a single part be revised without regenerating everything? Kimi's official October 30 tutorial acknowledged that its K3.1 Pro model "temporarily cannot do well" on mesh-based modeling for characters, animals, statues, monsters, and complex props, so it delegates such tasks to Tripo.

This reveals a shifting dynamic: the more LLMs push complex 3D projects, the more specific the demand for native 3D capabilities becomes. Agents can plan, write code, and check results, but they need accurate, fast, structurally clean 3D objects to continue editing, controlling, and organizing interaction. A new collaboration mode is forming between code intelligence and 3D native intelligence.

VAST's Two Research Lines: Tripo and Eden

VAST (三启万物) pursues two main research threads: native 3D generation and understanding centered on Tripo, and world-model exploration represented by Project Eden.

Tripo: Production-Ready 3D Generation for Agent Workflows

Tripo appears repeatedly in notable "LLM+3D" projects:

Independent developer Sac used Claude, Tripo, and Blender to turn The Little Prince 's planet into an explorable 3D scene in 16 days, earning 10k+ likes on X and shares from a16z and Marc Lou.

@thebuggeddev combined GPT, Codex, and Tripo to build a web-based 3D human anatomy app with interactive organ models, later retweeted by OpenAI's Greg Brockman.

At Anthropic's Claude Build Day, Holly reconstructed the five great Buddhas of Nanchan Temple using Tripo, winning first prize with a Tang-dynasty timber-architecture focus.

Shanghai Jiao Tong University's "SITU" cultural-heritage agent collaborated Tripo with Kimi to 3D-reconstruct the Silk Road across millennia.

These diverse projects share a creation path: LLMs handle planning, programming, and workflow orchestration; Tripo provides professional 3D generation; creators assemble them into experiencable works.

Benchmark Leadership and the P Series Philosophy

From March 2024 onward, Tripo models have topped a third-party leaderboard with 200k+ user votes for eight consecutive months. The H series pushes precision ceilings; the P series is the only 3D foundation model globally that outputs production-grade meshes in seconds , embodying a "generation equals usability" philosophy.

Tripo P1.0 launched in March 2024; P2.0 followed in September, generating quad-dominant meshes with controllable face counts and more logical part separation in seconds. These traits matter for developers and agents: speed enables multi-round "generate–run–check–modify" iteration; controllable complexity fits web, mobile, and real-time performance budgets; regular topology reduces downstream editing and animation friction; logical part separation lets code locate and manipulate object components.

VAST Chief Scientist Yanpei Cao demonstrated this by generating a mechanical spider with P2.0, then having an LLM name the natively formed parts, write and debug motion logic for each, achieving part-level animation control.

When agents already write code, the 3D native foundation model must supply 3D objects that code can understand and manipulate.

From Production Needs: Rebuilding Generation from the Ground Up

Post-processing cannot fix fundamental topology issues. Traditional AI-generated models often require manual retopology for editing, subdivision, and animation — time-consuming manually, unstable automatically. The P series' underlying Nexus framework moves the problem to the generation stage.

Unlike the industry-default approach of converting meshes to sequences for autoregressive generation, Nexus decouples vertex and topology generation: a diffusion model constructs vertices coarse-to-fine, while a novel encoding expresses topological relationships, enabling native mesh generation without pre-sorting. The paper was accepted at SIGGRAPH; a real-time demo based on P1.0 won SIGGRAPH 2026 Real-Time Live! Best in Show.

P2.0 adds native quad topology, eliminating the "generate then re-topologize" loop. The trade-off is clear: optimize simultaneously for shape fidelity, speed, geometric complexity, and downstream usability — not just visual detail or face count — so the output holds up in real production pipelines.

From Research Curiosity to Flagship Model

The P series began not as a planned product but as core contributor Ronnie's research interest during an internship. "There was no such target at the time, just interest in the problem, step-by-step exploration, and eventually this result." As experiments progressed, the line became a flagship model. Ronnie's growth included learning to communicate algorithmic capabilities, understand real user needs, and translate ability into usable experience. A pivotal moment came at GDC 2024 San Francisco when a game developer, seeing P1.0, said they had never seen such capability. Months later, P2.0's release sparked developers sharing novel workflows, integrating the model into agents Ronnie hadn't imagined.

Project Eden: World Models Beyond Pretty Pixels

If P series answers "how can a 3D object be better used downstream," Project Eden asks: when these objects share an interactive world, how does that world persist and evolve?

Action-conditioned video models like Google Genie 3 predict "next pixels" from short history; once an object leaves the frame, it may vanish on return. In June 2024, Fei-Fei Li's World Labs published "A Functional Taxonomy of World Models," proposing a planner–simulator–renderer framework. Four days earlier, VAST unveiled Project Eden with a similar decoupled three-layer architecture:

Persistent world state: records object positions, attributes, interaction rules, action consequences — independent of any camera view.

State-to-frame interface: converts current state into layout, geometry, and event constraints for a given viewpoint.

Generative renderer: produces high-fidelity frames from those constraints.

Since August 2024, numerous works (Code World Model, Magpie, Programmable World Model, GameDirector, CoDeR, AgentGarten, Matrix-Game 4.0, Streaming Code World Model, SolarRenderer, NVIDIA StreamRender-H3) adopt this paradigm: code, game engines, or agents maintain world state; video models render frames.

EdenLive: Open-Source Real-Time Generative Renderer

On October 18, 2024, VAST open-sourced EdenLive — the first video model that generates different-styled frames in real time from arbitrary visual references. On a single GPU, it exceeds 24 FPS at 720p, supports minute-long continuous generation; the 5B model runs in real time on a single RTX 5090.

Core contributor Zion (a 2000s-born researcher) joined VAST as an intern before grad school, later interned at a big-tech LLM team with a return offer, but chose VAST for campus recruitment. When world-model hype favored end-to-end action-conditioned video, VAST's team asked: "What is a world model conditioned on keyboard/mouse actually good for?" They concluded explicit world state is essential — games maintain health bars, cooldowns, etc., but the prevailing paradigm lacked this. The decoupled approach faced internal challenges: "It differs from the mainstream; end-to-end looks more technically impressive. Does decoupling avoid the harder problems?" The team returned to first principles: they wanted a persistently interactive world with clear rules, not just a responsive video. Months later, industry direction validated this judgment.

Engineering Persistence: From Drift to Stable Streaming

Zion, previously a 3D generation researcher, built the low-level R&D framework to plug in new video backbones and rapidly test streaming training and distillation methods. Early experiments suffered frame drift around the 30-second mark. He debugged numerical precision and implementation details, combined updated algorithms, and iteratively achieved stable quality. This framework underpinned early Eden experiments and EdenLive iterations. The company provided data, infrastructure, and compute; the team discussed next-step hypotheses collaboratively. For Zion, the key growth was clarifying long-term research goals: interaction and simulation — how 3D content enables new interaction experiences, how to efficiently build simulation environments for agents and robots — linking 3D generation, coding agents, and world models concretely.

Culture: Freedom to Define Problems, Support to Solve Them

Both research lines share a working style: decide what problem is worth solving, then find the method; dissent is welcome; experiments can overturn assumptions. "Research freedom" means the freedom to define the problem itself .

Ronnie values that the team respects facts and the research process: discussions focus on how to solve, not on hitting pre-set breakthrough dates.

Zion's early attempt to use LLM/VLM agents with 3D generation for scene layout stalled because base models couldn't reliably correct object orientation. The project was paused after limited-resource validation failed — not a rejection of the researcher, but a staged mechanism: test key assumptions with minimal investment; continue if promising, redefine if not.

PhD intern Stewie participated in the full core-model pipeline from data to training — beyond his lab experience of chasing paper metrics. He now balances user needs, stability, and qualitative experience details. His proudest moments come when a dataset or method solves a product problem and improves user results, not from publishing another paper.

Algorithm member R likens the team to a scientific expedition: facing the unknown, members need direction judgment, product understanding, and cross-disciplinary collaboration — not just a pre-drawn map. He seeks "high capability, low ego" partners. Stewie experiences this as daily support: colleagues share understanding and technical experience across specialties, dissect problems together, letting him focus on the problem rather than impostor syndrome. High talent density yields support, not peer pressure. "No internal friction" means direct disagreement, cross-team implementation discussions, and owning problems without boundary negotiations.