In the gaming industry, there’s a type of work that’s rarely discussed but consumes a massive amount of time and budget. It’s not gameplay, it’s not level design, and it’s not the art style that people talk about more often. Rather, it’s the tasks in the latter stages of the art pipeline: retopology, UV unwrapping, weight painting, and rigging…No matter how beautifully a character’s high-poly model is sculpted, if the underlying topology is flawed, it will become obvious as soon as animation is applied. What follows is a lengthy process of backtracking: fixing the topology, re-UVing, re-texturing, re-rigging, and re-animating—a domino effect where changing one thing affects everything else. While these tasks often involve little creativity, they are precisely what determine whether a 3D asset can truly enter production.Over the past two years, AI-generated images and videos have raised everyone’s expectations sky-high, so AI 3D has naturally become the next big thing. Many people are working on it, and there are plenty of impressive demos, but those who actually need to use these models in production know full well that there’s still a long way to go from “usable” to truly “effective” and “user-friendly.”Many high-poly models look impressive at first glance, but they fail the very first test when integrated into the pipeline. The underlying mesh is a jumble of illogical, fragmented faces that can’t be edited, can’t handle real-time rendering, and make animation practically impossible.What exactly is the difference between “3D that looks like 3D” and “truly usable 3D”? Is anyone actually bridging this gap? In this episode of the 405 Gaming Podcast, we’re joined by Cao Yanpei, Chief Scientist at VAST.Founded in 2023, VAST focuses on AI 3D large models and world models. The company recently completed a $50 million Series A funding round co-led by Alibaba and Hengxu Capital. Its product, Tripo, has already seen numerous practical applications in the industry.Tripo P1.0, released this past March, focuses not on “better visuals” but on “native usability,” aiming to directly output standard low-poly meshes with clean topology that are ready for engine integration within 2 seconds.
Mere functionality is just the first step; assets must also be dynamic and interactive. The next challenge is the automatic generation of binding and animations, followed by endowing the assets themselves with behavioral logic. Following this path leads to world modeling: not just creating objects, but building a world in which those objects can function.Recently, they also announced early research in the field of world models, and it’s said that their first world model is coming soon. Cao Yanpei is the technical lead behind these models; he holds a bachelor’s and Ph.D. from Tsinghua University and previously led generative 3D research at Tencent.During a conversation lasting nearly an hour and a half, he used the metaphor of a “paper lantern” to clearly explain the composition of 3D assets: what meshes, topology, UVs, and skeletons each do, and why rework is constant in traditional pipelines.He also didn’t shy away from acknowledging current shortcomings: while topology issues have largely been resolved, rigging and animation are still some way off from being production-ready.However, precisely because the path of 3D is indeed more challenging, his judgment regarding the choice of approach is all the more noteworthy: while most researchers are betting on video as the primary path to world models, he believes this is treating a dimension-reduced projection as a native signal.3D, on the other hand, is the natural representation of the physical world, and its long-term value has been severely underestimated. Beyond the technical route itself, he also raised a broader question: If you could generate 100,000 assets in a single day, would you still make games the way you do today?
Perhaps once the barrier to entry for 3D creation has truly dropped low enough, we’ll see some new developments in the interactive entertainment sector.
The following is the complete transcript of the conversation, compiled by Game Teahouse with authorization; the content has been edited.
01
From “looking like” to “actually working”—what exactly makes 3D so challenging?
“Images and videos are just a 2D layer of pixels, whereas 3D models must have precise structures, reasonable topology, and a skeleton system that allows for physics-based collision calculations within the engine—there’s a huge gap between the two.”
Xiaoning: Over the past two years, everyone has seen a lot of AI-generated images and videos, but if we take it a step further, 3D is actually a more critical—and more challenging—topic—it’s not just about generating something, but about truly integrating it into the game development and content production processes.
Today, we’d like to ask more specific questions: What exactly makes AI 3D so challenging? What steps must a pipeline go through to be ready for production? And where does the industry stand today?
Cao Yanpei: You’ve pinpointed the pain point in the field of 3D generation very accurately. Over the past year or two, whether it’s Sora, Seedance, or various other AI tools for image generation and editing, people have been “spoiled” by the stunning visual effects—so isn’t it only natural to expect the same from 3D generation?But anyone working in gaming or 3D interaction knows there’s a huge gap between something that “looks like 3D” and “3D assets that can actually be used in a game engine.”Images and videos are merely a “skin” of 2D pixels, whereas 3D models must have precise structures, topological wiring, skeletons, and the ability to handle physics-based collisions within the engine, among other requirements.
At Tripo, we’ve focused on one thing over the past few years: how to ensure that AI-generated content is no longer just a “pretty but useless” visual illusion, but rather industrial-grade assets that can truly help artists and developers generate animations and interactions in just seconds—ready for use in game engines or DCC software like Maya.We recently released Tripo H3.1 and P1.0, which feature a paradigm-shifting overhaul of our underlying architecture. We’re delighted to take this opportunity to delve deeper with you into the technical logic behind AI 3D.
Xiaoning: Many of our listeners are game industry professionals, but even among them, there are probably more game designers and programmers than people who truly understand the 3D pipeline.Let’s break this down step by step. From a fundamental perspective, what is the most fundamental difference between an image, a video, and a truly usable 3D asset?
Cao Yanpei: The most fundamental difference lies in the inherently different nature of data representation. When generating images or videos, neural networks process extremely regular, dense pixel matrices—the AI simply needs to fill in the color of each cell within a fixed grid.However, 3D is, by its very nature, a sparse structure in space—or even a non-Euclidean space. AI must not only determine the positions of tens of thousands of vertices but also construct the connections between them—that is, the “mesh topology.”
2D modeling studies the statistical laws of light and shadow on a plane; while it can now produce highly realistic lighting effects, it is essentially a form of “visual illusion” with no actual three-dimensional entity behind it. For an object to exist in three-dimensional space, its surface must be represented continuously and accurately, and it must also possess a complete structure and functionality.Making the leap from fitting 2D lighting patterns to constructing rigorous physical and topological logic is not merely a matter of mathematical principles; it requires more careful consideration in the design of neural networks and generative models.
Xiaoning: Meshes, topology, UVs, and skeleton binding—could you explain why they’re important in the simplest, least technical terms?
Cao Yanpei: Imagine a 3D model as a paper lantern. The mesh is the shape formed by the pieces of paper that make up the lantern’s outer shell.Topology refers to how the paper pieces are joined together; in the human body, this corresponds to the direction of muscle fibers. Why is topology so difficult, and why does it have such a significant impact? Because if the topology is poor, the model cannot move—just as muscles growing haphazardly would cause a bent arm to tear—and when animating, the model would look like a crumpled ball of paper. UV mapping is the process of unfolding a 3D surface onto a 2D plane, much like peeling an orange—if the UVs are too fragmented, the texture artists will be driven to despair, as the surface becomes disjointed and the edges will never align properly.Skeletal rigging is like stuffing a movable metal skeleton inside a paper lantern, specifying which bones pull on which pieces of paper to make the lantern move.
Xiaoning: That’s quite a vivid description. Setting AI aside for a moment, how does the traditional 3D pipeline typically work? Which steps take the most time?
Cao Yanpei: Take the next-generation game development process as an example: First, there are the concept art designs, then the sculptors and modelers create an extremely detailed “high-poly model” with a very high polygon count.Next, retopology is performed, where the high-poly model is covered with a simplified, rule-based polygon mesh to create a “low-poly model.” Once the low-poly model is ready, UV unwrapping and texture painting are done, followed by the application of a skeleton and weighting. Finally, the model is handed off to the animators.
The most time-consuming tasks are retopology and rigging; they involve very little artistic creativity and are essentially tedious manual labor. Moreover, rigging and animation are areas where painful rework often occurs—many times, it’s only after handing the assets over to animators to bring the character to life that we discover issues with the topology and mesh layout when the character bends over or throws a punch.Since the entire workflow is linear, changing the topology requires re-UV mapping, re-texturing, re-rigging, and re-animating—a single change affects everything else.
02
Don't Let Others Clean Up After AI: The Entire Process from Model Generation to Production
“The underlying philosophy behind P 1.0 is that, at this stage, AI needs to adapt to the industrial standards that humans have developed over decades, rather than having humans clean up after models generated by AI.”
Xiaoning: A couple of years ago, when people were talking about AI 3D, I got the impression that everyone was working on what you called the “first step”—generating high-quality 3D models. But based on your explanation just now, it seems like people haven’t yet delved into the real core issue?
Cao Yanpei: You’re absolutely right. Our primary focus is on the issue of mesh topology, which has been overlooked by many in the past couple of years. Laypeople only care about whether a model looks realistic, but when modelers or technical artists in the industry receive a model, the first thing they’ll likely do is press a shortcut key to switch to wireframe mode and check if the underlying wireframe is correct.If the AI outputs a dense, illogical cluster of triangles, it creates a very heavy asset that poses significant performance challenges for both editing and real-time rendering. It’s also difficult to use directly during the animation phase, and such high-poly models don’t fit into existing industrial pipelines.
Our P1.0 aims to address the challenge of generating low-poly models that are ready for pipeline use—what we call “natively usable” assets. Today, we believe it is indeed time for AI 3D to move beyond “generating high-poly models solely for proof-of-concept” and enter a “true era of productivity.”
The underlying philosophy behind P1.0 is this: at this stage, AI needs to adapt to the industrial standards that humans have developed over decades, rather than having humans clean up after AI-generated models. The moment a model is generated, it is already a standard mesh—primarily composed of quadrangles, with proper wiring and accurate geometry—ready to use straight out of the renderer.From high-poly models to topology, and on to UVs and textures, these assets can essentially be imported directly into the UE engine or Maya for animation and simulation.
Xiaoning: “Native support” means it can be edited at several key stages and can be fed directly into the production pipeline?
Cao Yanpei: Yes, it’s more about integrating directly into the workflow without requiring a lot of manual corrections.If it takes AI two or even ten minutes to generate a model, but because the topology is a mess, a highly skilled artist still has to spend half a day cleaning up vertices and retopologizing it, then it becomes very difficult to determine whether AI is actually a tool for improving efficiency or one that increases rework costs.
Xiaoning: I’ve heard of two approaches to addressing 3D workflows: one involves tackling each issue individually, while the other is more like addressing all stages at once. What are the respective advantages of these two approaches?
Cao Yanpei: My view is this: At the individual task level, it’s definitely a sudden change—one day, a certain technology suddenly becomes usable; but its integration into the entire pipeline is gradual. For example, mechanical tasks like manual retopology and manual UV unwrapping—P1.0 is already very close to a breakthrough point, and these tasks may soon be taken over by AI.However, for the foreseeable future, generated models will still need to flow between standard workflows in standard formats. We don’t want to completely change how people use engines; rather, we aim to replace the way assets are produced at each stage. While technological breakthroughs are sudden, when each breakthrough is integrated into production, we still hope to incorporate it into the pipeline gradually.
Xiaoning: So how will this gradual integration take place? How will AI and humans collaborate?
Cao Yanpei: We’re approaching this from two angles. First, regarding the professional workflow, we hope to use AI to replace the most tedious and least creative aspects—such as topology and binding—as thoroughly as possible, so that everyone can focus their energy on the parts that truly require creativity.
On the other hand, the greatest potential for AI 3D technology may lie with the general public—people who have never worked with a 3D pipeline before or even know what a 3D model is. Once AI can cover the entire pipeline, end users won’t need any understanding of computer graphics or the production process for art assets.Just as when you download an icon from the internet, you don’t care whether it was created with AI or Photoshop, nor do you care how an SVG animation was made.3D AI must also bring this ultra-low barrier to entry to ordinary users through end-to-end capabilities, so that people no longer view 3D as a mysterious or niche form of expression—only then can a broader UGC ecosystem emerge.
Xiaoning: Today, the groups that actually use 3D assets are still animation studios, game developers, and indie developers. What kinds of people have embraced AI-powered 3D? And what kinds of companies are still relatively conservative?
Cao Yanpei: Right now, the teams that stand to gain the most value are definitely small and medium-sized—or even independent—game development teams, as well as those working on UGC platforms or procedural generation. Large AAA teams have massive art budgets and tend to be cautious about changing their pipelines. But for small and medium-sized teams, “ideas often outweigh production capacity.”A lack of art resources is a major bottleneck limiting their rapid development and the realization of their core creative visions. AI 3D provides them with a powerful lever, allowing them to build worlds—which would otherwise require a team of dozens of artists—at an extremely low cost.
As for P1.0, it has essentially reached a level where it can be used out of the box for environmental props, mid-to-long-range architectural components, furniture, and even simple NPC characters and vehicles.The next step is to tackle highly complex animations—especially facial animations—as well as game protagonists with rigged animations. If we can overcome this hurdle, acceptance among AAA studios will gradually increase. After two or three years of iterating on 3D generation technology, we’ve definitely seen that as quality improves, more and more upstream users are recognizing its value.
Xiaoning: When I used to work on games, I had this feeling: if the total art assets were 100, the top 20%—characters, CG, and core animations—might account for 80% of the budget, while the remaining 80% of the “filler” only consumed 20%.If you don’t master the most critical aspects, won’t you miss out on the lion’s share?
Cao Yanpei: It’s certainly crucial for AAA productions, but we’ve always placed equal importance on lowering the barrier to entry.At this very moment, there are very few people in the world who are truly capable of producing games—they represent the very tip of the pyramid. To bring more creators into the field, we need to expand the pool of people who have access to this technology as widely as possible. In addition to continuously refining the most intricate aspects of generation, we also need to ensure that, in terms of usability and speed, everyone perceives the creation of 3D assets as something that requires no barriers to entry and incurs no cost.
Xiaoning: Let’s first discuss the professional standards in more detail. Core assets aren’t limited to 3A titles; the main characters, facial models, and animations in many second-tier or relatively large-scale games also qualify as core assets. What’s still missing to break through this ceiling?
Cao Yanpei: P1.0 already covers the early to mid-stages of the process—from high-poly to low-poly models, topology, and initial texturing.The most challenging part—and the one we’re devoting a lot of effort to solving—is dynamic binding. AI must not only understand geometric shapes but also the mechanical structures and physical motion mechanisms behind 3D objects. This involves processing a much broader range of information, and accumulating and cleaning the data is also more time-consuming.
The key is that it’s not just about the accuracy of the assets themselves—all components need to be complete. Users should be able to start working with them and creating animations right away. While the binding and animation features may not yet be fully “production-ready,” I believe we’ll see breakthroughs in this area very soon.
Xiaoning: More than one company is working on this. If they actually succeed, will it become a general-purpose technology? What exactly are they competing over?
Cao Yanpei: Given the current stage of the industry, AI 3D is shifting from the proliferation of purely algorithmic research papers seen in the past two years toward a phase of defining industrial data standards. Whoever can provide the most stable, standardized, fast, and easily integrated data format for existing engines will be most likely to secure a commanding position in the ecosystem.
This involves the classic “innovator’s dilemma.”Major companies with massive game portfolios have excellent technical teams, computing power, and funding, but their in-house AI teams are burdened by a heavy historical legacy—they must adapt to existing, even outdated, workflows, such as using AI developed a decade later to support game production processes from a decade ago. The tight coupling between specific workflows and business logic traps R&D in a dilemma.
Independent platforms like VAST, on the other hand, offer the advantages of technology neutrality and agility—they don’t patch specific legacy games or cater to any particular production workflow. Instead, they can start from first principles and use cutting-edge algorithms to directly define what the future of 3D generation should look like.
In the longer term, the core challenge we’ve consistently been tackling is: how to instantly and directly generate an interactive 3D world with underlying logic.A perfect 3D asset should have skin, flesh, bones, and a brain. In the past, most models generated by the industry only achieved the “skin”—they had basic visual elements, but everything beneath that was missing. P1.0 solved the “flesh” issue through native meshes—a seamless topological structure.Animation rigging represents the “bones”—endowing the model with kinematic and physical structures, placing joints in logical positions, and enabling it to respond to user interactions. The “brain” refers to the asset itself potentially functioning as an NPC or agent, interacting with players and other agents within the environment.
Progressing step by step from the generation of pure geometric shapes to the construction of assets that incorporate kinematics and interactive logic has been our technical goal from the very beginning.
Xiaoning: Based on what you just said, the entire pipeline workflow will change. A lot of manpower has been invested in the 3D pipeline, so as professionals in the field, we naturally wonder: What won’t be replaced? What’s likely to be replaced?
Cao Yanpei: With AI developing so rapidly, it’s not just artists and game developers who feel a sense of crisis—even programmers and researchers, whom people used to consider at the pinnacle of intellectual achievement, are now deeply concerned.This topic requires long-term discussion. But returning to the 3D pipeline itself, the significance and trends of technological development—which also align with the direction we hope to take in building technology—are to free everyone from tedious, repetitive tasks so they can channel their energy into areas that truly require creativity and innovation.Even when limited to the 3D pipeline or the gaming industry, what will truly demonstrate one’s value and set one apart in the future will still be ideas, creativity, and the ability to take action.
03
Two seconds after the model is generated, when the speed is fast enough to change the way you create
“If someone told you that you could generate 100,000 assets in a single day, what kind of game would you design? Compared to a scenario where it takes half a month to create just one main character asset, people would make very different choices. That first option simply didn’t exist before.”
Xiaoning: You’ve mentioned speed several times. Why do you place such an emphasis on it?
Cao Yanpei: For us, speed isn’t just a selling point—it’s been a goal we’ve pursued since the company’s inception. A quantitative increase in speed can trigger a qualitative leap in 3D content creation. The most significant benefit right now is that it reduces the cost of trial and error to virtually zero. Generating a model takes 10 minutes, which interrupts a creator’s train of thought. But P1.0’s 2-second generation speed enables real-time feedback—you can instantly test 10 different concepts and keep the best one.
This actually frees up creators’ imaginations—there’s no longer much resistance to whatever comes to mind. It’s just like how people write code today: when an idea strikes, they immediately feed it into Codex or Claude, then check the output to see if it’s correct. The same goes for 3D: only by infinitely reducing the generation time—and the psychological pressure that comes with it—can we hope to spark a true qualitative leap.
Speed is even more critical for the UGC ecosystem. While speed leads to increased efficiency in professional development, in UGC, it provides the most essential sense of instant gratification. Ordinary users don’t have the patience to wait 10 minutes for a progress bar to fill; only AI can instantly generate 3D objects—just like hitting the Enter key—which gives users the motivation to continue interacting and creating.
Xiaoning: I think when it comes to speed, whether you view it from the perspective of production or consumption, your expectations are completely different. I really agree with that logic regarding UGC. But when it comes to the production side, creating a 3D asset used to take days or weeks—reducing that to 2 seconds is extreme.At what speed does a qualitative change actually occur? Is such an extreme reduction really necessary?
Cao Yanpei: The 2-second process primarily focuses on generating native mesh topology; subsequent steps—such as applying textures and materials, and rigging animations—still require manual intervention or collaboration between AI and humans.It’s not an end-to-end process that generates a complete asset in a matter of seconds, but that’s a goal we’re very eager to pursue next. Do we really need it to be in the seconds range? Perhaps not, but the faster the generation, the more people will be able to break free from the mental constraints imposed by previous time limitations.
If someone told you that you could generate 100,000 assets in a single day, what kind of game would you design? Compared to a situation where it takes half a month just to create a single main character asset, people would make very different choices—the former option simply didn’t exist before. Even in specialized fields, quantitative changes in speed can still lead to qualitative changes in production paradigms and approaches.
Xiaoning: As someone who used to work in the gaming industry, this is definitely a positive development for creative developers. But for regular workers, it seems like the competition has gotten even fiercer—whereas before we might have created one or two assets a week, with these tools, we might have to produce 10,000 a day?
Cao Yanpei: I don’t necessarily think it’s more competitive. With AI capabilities, the format of content creation will also change.It’s a bit like comparing the people who used to produce videos at TV stations to the professionals creating short videos at MCNs today—while their core personal value may be the same, the formats of content have become more diverse. In addition to long-form content, there are now short dramas and an endless stream of short videos, which truly offer more room for talent to shine.As for whether competition has intensified, that’s a complex social and economic issue that cannot be attributed entirely to a single technological advancement.
Xiaoning: When you talk about UGC, are you more concerned with lowering the barrier to entry for developers, or do you really want to make it accessible to everyone right from the start?
Cao Yanpei: The ultimate goal is a more pure form of 3D UGC—that’s why we started this company in the first place. But every technology that lowers the barrier to entry also reduces costs and improves efficiency to some extent; you could call it “picking up bonuses along the way.”Most people haven’t clearly defined what 3D UGC actually looks like, and many are even skeptical about whether it can be achieved at all. So we always aim for lofty goals, but we also make sure each technology plays its proper role in today’s industry in a down-to-earth way.
When we introduce Tripo, we don’t usually say, “We’re a 3D tools company that helps game companies save time and effort.” If we position ourselves purely as a tools provider, our value boils down to cost reduction and efficiency gains. What we’re actually doing is unlocking the underlying infrastructure for the next generation of UGC or mass-market interactive entertainment platforms.
Xiaoning: Is there a priority order between lowering the barrier to entry and pushing the upper limits? Do they involve the same technical challenges?
Cao Yanpei: Under today’s industry conditions, it’s essentially the same thing. Even as the entry barrier is lowered, assets must still be “pipeline-ready.” While striving for higher quality, they must still be pipeline-ready assets—once they’re pipeline-ready, the requirements for assets are the same, whether they’re for game engines or Vibe-coded mini-games.With the ability to generate pipeline-ready assets and an increasingly comprehensive set of components, the barrier to entry naturally lowers. At least for now, these two are one and the same.
04
Betting on 3D Instead of Video: A Different Bet on the Future
“The real world is inherently 3D; the 2D pixels in videos are merely projections resulting from the compression and dimensional reduction of the three-dimensional world. The long-term technical scalability of the native 3D approach has been greatly underestimated in the field of AI research.”
Xiaoning: You’ve been talking about game engines, but in this era of video model development, people are also wondering if game engines themselves will change. Is there a possibility that, in the future, certain game genres will emerge that don’t rely on engines at all?
Cao Yanpei: It’s entirely possible; we’ve spent a lot of time thinking about and exploring this ourselves. Looking ahead to the next few years, interactive content won’t be limited to traditional game formats. For example, there are already interactive world models where you start with a single scene and, through video models and on-screen interactions, explore the environment and even influence it.
Also, as AI coding and agent capabilities improve, is there a need for a new engine that’s easier for AI agents to use? This is likely to be a future trend. Currently, vibe-coded games all use Three.js; from my perspective, this is because AI doesn’t yet have the capability to effectively handle heavy engines like Unity or UE, so the most straightforward approach is to write JavaScript from scratch.
Xiaoning: The term “world model” has become increasingly popular over the past six months and has sparked a lot of discussion. How would you define a world model? How does yours differ from others?
Cao Yanpei: I agree with the commonly accepted definition in academia: given a current state and an action, predict the state of the environment at the next time step. However, people apply this concept in their own respective fields of interest.
We take a different approach from the most popular pure video generation methods in the industry. Pure video generation is more of a lighting simulator; it focuses on learning how light and shadow change, but it’s difficult to ensure the underlying rules of the 3D world are accurately represented. This is also why video-generated worlds are often criticized for containing common-sense or logical errors.
From the very beginning, our goal has been to enable video models to function in an environment characterized by concurrent multi-player interaction and highly controllable mechanics and behaviors—one that supports multiple players and agents simultaneously influencing the environment.More importantly, the mechanics and behaviors within the world must be highly predictable and deterministic—in a racing game, if Player A overtakes Player B, Player A is ahead of Player B; this cannot change simply because their positions shift for a single frame. From a fundamental mechanical standpoint, pure video models find it difficult to guarantee this uniqueness and determinism.
Of course, some might argue that the video world model—where “you don’t know what a place will look like when you return to it”—offers a novel experience. However, we believe this sense of novelty is essentially a stopgap measure built on the memory limitations of video models, and it does little to help creators who truly want to take control and be creative in building content that can be experienced repeatedly.Our approach to world modeling remains the same: it supports multiplayer interaction, and the world’s mechanics and future are predictable and uniquely determinate.
Xiaoning: The logic behind pure video generation isn’t true understanding in the strictest sense, and small errors can accumulate. To achieve the goal you mentioned, it seems to imply that the model’s reasoning must involve a genuine logical sequence?
Cao Yanpei: Absolutely correct. That’s why we need to move beyond the paradigm of standalone video models and entrust both inference and visual rendering to a single model; with our current capabilities, it’s difficult to ensure a high degree of controllability.We will introduce more explicit states—such as 3D or code-based states—during the inference stage. This is still a fairly early-stage exploration, but once the path forward is clear, we will iterate and validate it quickly.
Xiaoning: Why not just wait for someone else to develop it and use it?
Cao Yanpei: Because opinions vary widely on what the world model should look like and which problems to tackle first, the choices in technology differ significantly. The mainstream application scenarios are likely to be robot control and simulation, which are quite different from the applications we’re focusing on for the next three to five years; this will lead to the emergence of different technical approaches.Furthermore, Tripo has always focused on addressing “states” within the digital world—generating assets is essentially generating states, but how those assets evolve and how states transition from one to another are precisely the challenges that world models are designed to solve. We have arrived at this critical juncture, and world modeling is a technology direction with immense potential and scope for innovation.
Xiaoning: Where will your most critical investments be focused over the next one to two years?
Cao Yanpei: In the field of AI, a two- or three-year plan is really quite long—the underlying mechanisms of algorithms, computing power, neural network architecture, and multimodal research paradigms may undergo major restructuring every few months.For us, the key is to maintain a high level of technological acumen and agile decision-making capabilities, avoid becoming overly reliant on any single approach, and be able to adjust our direction very quickly when new trends emerge—following a highly agile iterative path from technological restructuring to product-level models and ultimately to product capabilities.
I envision that in two or three years, we’ll have completed the transition from a 3D asset productivity tool to a platform for spatial computing or UGC interaction—but in reality, everyone is more focused on rapid iterations over the next two or three months, or even just a few weeks.
Xiaoning: How long will it take for AI 3D to be integrated directly into the production pipeline at a high quality level? Will we see this happen in two or three years?
Cao Yanpei: Given the current pace of technological advancement, two or three years might even be a bit too long. By this time next year, we can fully expect AI-generated assets to feature high-quality texture mapping and topology, making them suitable for use as fully functional game assets.
This time last year, there was a lot of criticism about the quality of high-poly models themselves—even a randomly generated car wouldn’t have the right shape, and the figures weren’t up to the standard needed for 3D printing.But a little over a year later, high-resolution models have nearly reached the level of hand-sculpted work by human artists, and topologically sound low-poly models can also be produced very well. By next year, assets that are “fleshed out—with skin, flesh, bones, and a brain”—should be able to be generated end-to-end by AI.
Xiaoning: From today’s perspective, what aspects of AI and 3D are overrated? What aspects are underrated?
Cao Yanpei: What’s still overrated is visual realism. People see very lifelike 3D-generated results online, but those are presented in video format. From the perspective of the rendered high-poly models, they are indeed detailed—industry professionals all share this impression, while the general public can’t really tell the difference.
What is underestimated is the ultimate value of 3D signals as a native representation of the physical world. In the research community, the 3D modality is largely viewed with skepticism.Many scholars cite “bitter lessons” to argue that 3D is difficult to scale, lacks generality, and involves a great deal of manual design, believing that the winning approach must be one that relies more purely on massive amounts of video data. I believe this perspective has significant limitations—the framework it implies is constrained by the mindset of traditional computer vision tasks.
Let’s go back to first principles: Which is more native—a 2D pixel matrix or a 3D signal with absolute spatial dimensions? The answer is clearly the latter. The real world itself is 3D; 2D pixels are projections resulting from the dimensionality reduction of the three-dimensional world.P1.0 learns and generates directly within a three-dimensional structure; rather than introducing artificial designs, it allows AI to directly fit the original signals of the physical world. By adhering to the native 3D approach, its long-term technical scalability and sustainability have been greatly underestimated.
Xiaoning: During GDC, Simon (founder of VAST) told me that many people haven’t really experienced AI 3D yet, even at a major developer conference like GDC. Could this also be a form of underestimation?
Cao Yanpei: Personally, I feel exactly the opposite—I’m quite optimistic. The fact that people haven’t yet had the chance to experience something that’s clearly useful means there’s still enormous potential for the future. It’s easy to understand why this is the case: it wasn’t until P1.0 was released that we were able to quickly obtain assets with correct topology.Earlier, the first group of users who tried it found it was high-poly and didn’t see much value in it, so they had no incentive to promote it. After P1.0 was released, everyone realized that what was being delivered were genuine game asset meshes, and many overseas clients proactively reached out to integrate the API; after trying it out, they naturally signed up for subscriptions. This was indeed a major turning point.
Xiaoning: If you had to describe Tripo in just a sentence or two, how would you like the public to understand it?
Cao Yanpei: We hope everyone will view Tripo as the foundational platform for a brand-new UGC interactive platform and 3D content ecosystem of the future. It’s not just a time- and effort-saving 3D creation tool—it’s a complete set of foundational capabilities designed for the next generation of interactive formats and the 3D content ecosystem.Whether you’re part of a large AAA development team or just an ordinary person with no artistic background—but plenty of passion and a head full of ideas—you can build the 3D worlds in your imagination in real time with a very low barrier to entry.
Xiaoning: This is a truly grand vision.
Cao Yanpei: I won’t deny that this is a very ambitious goal. But from the very beginning, our team has always wanted to keep iterating while aiming for ambitious goals, rather than limiting ourselves to a future we can see coming too soon.
原创文章，作者：游茶妹儿，禁止转载：https://youxichaguan.com/en/archives/216294
