How Hyper3D Rodin Gen-2.5 Is Bringing Production-Level Control to AI 3D Generation

Over the past several years, AI-generated 3D content has rapidly evolved from experimental demos into practical tools increasingly used across game development, VFX, XR, product visualization, and real-time production workflows. Yet despite dramatic improvements in generation quality and speed, many artists and studios still struggle to integrate AI-generated assets into real production pipelines where topology, controllability, optimization, and downstream compatibility matter just as much as visual fidelity. 

That production-focused gap is exactly where Hyper3D positions its latest update to Rodin Gen-2.5. Rather than approaching AI purely as a one-click generation tool, the company describes Rodin as a controllable AI-to-3D workflow system designed to collaborate with artists across different stages of production.

In this interview, QX Zhang, CTO of Hyper3D discusses why “controllability” has become one of the most important challenges facing AI 3D workflows, the technical barriers separating prompt-generated assets from production-ready content, and how Rodin Gen-2.5’s new Adaptive Thinking Effort system allows artists to balance generation speed against fidelity depending on their specific workflow needs.

To start, could you briefly introduce Hyper3D and Rodin Gen-2.5, and share how Hyper3D Rodin has evolved into one of the leading AI 3D generation platforms used by creators and enterprise teams? 

Hyper3D is a 3D creation platform we launched in 2023, built specifically for game developers and 3D creators. It offers AI-powered image-to-3D and text-to-3D asset generation through Rodin, character creation through ChatAvatar, and a suite of commonly used 3D tools under OmniCraft. More recently, we have also integrated new Image and Video foundation models into the platform to further expand its creative capabilities.

At the core of the platform is Rodin — our generative 3D foundation model, named after the renowned sculptor Auguste Rodin — which has now evolved into Rodin Gen-2.5. Rodin was among the earliest systems to bring truly native 3D generation into the industry, and the research behind it has received multiple SIGGRAPH Best Paper awards and nominations, including:

• SIGGRAPH 2024 Best Paper Nomination — CLAY
• SIGGRAPH 2025 Best Paper — CAST
• SIGGRAPH 2025 Top 10 Paper Fast-Forward — BANG

From the very beginning, we have envisioned Rodin as a tool that collaborates with artists rather than replacing them. “Controllability” has always been at the center of its design philosophy. Beyond basic Image/Text-to-3D generation, Rodin also provides uniquely creative capabilities such as 3D ControlNet, localized 3D Editing, part-level refinement, and production-friendly topology options.

In Gen-2.5, even generation speed and geometric detail density have become controllable parameters, which makes Rodin more adaptable for different workflows, from fast game asset ideation to high-detail VFX and 3D printing use cases.

That is also why Rodin has resonated so strongly with creators and artists around the world.

You can explore Rodin Gen-2.5 yourself with a 14-day free trial exclusively for 80 Level readers — use code 80lvRodinGen25.

A lot of AI 3D tools today focus primarily on fast generation. From your perspective, what does “controllability” actually mean in AI 3D, and why is it becoming increasingly important for production teams?

For a long time, generative AI was treated like a slot machine — sometimes you get the perfect result instantly, and sometimes you spend all day rerolling outputs.

That randomness made AI difficult to integrate into real production workflows, especially for game development, VFX, real-time rendering, and technical art teams where shape, topology, texture quality, and iteration control all matter. It also made AI feel opposed to artists — because no artist wants an uncontrollable tool.

From day one, “controllability” has been Rodin’s core philosophy. We see Rodin not as a one-click generator, but as a creative tool that collaborates with artists.

In early versions of Rodin, we introduced 3D ControlNet, allowing users to guide generation using Bounding Boxes, Voxels, and Point Clouds for precise shape and proportion control. In Gen-2, we introduced BANG-based part decomposition for flexible part-level refinement. In Gen-2 Edit, artists could upload existing meshes and use natural language for localized editing. And in Gen-2.5, even generation speed and geometry detail became controllable.

We don’t think speed and controllability should conflict. A good AI tool should let artists decide when they want fast iteration, and when they want to spend more time pursuing higher fidelity and detail.

One of the biggest criticisms developers often have about AI-generated assets is that they can look impressive in isolation but break down once brought into a real production pipeline. What are the biggest technical barriers between “prompt-to-3D” generation and truly production-ready assets?

Unlike text, images, video, or audio — which are mostly consumer-facing and already have relatively standardized formats — 3D is an entire industry, and different industries have completely different requirements for 3D assets.

A game-ready asset, a 3D printing model, a VFX hero asset, a real-time XR object, and an e-commerce product model may all need different geometry density, topology, UVs, texture maps, file formats, and optimization standards. 

Because of that, there is no single “general” 3D representation that works for everyone.

This is why AI 3D generation today is best understood as part of a workflow: it reduces workload rather than replacing the entire pipeline. For AI-generated 3D assets to become truly production-ready, they need to work with artists’ existing tools and pipelines, including DCC software, real-time engines, mesh optimization, PBR texturing, and downstream editing. 

Rodin Gen-2.5 introduces what you describe as “Adaptive Thinking Effort,” allowing generation modes that range from ultra-fast outputs to high-detail hero assets. How does that system work under the hood, and how should artists think about balancing speed versus fidelity?

It works in a similar spirit to the “thinking” mechanism in large language models, where the system allocates different levels of computation depending on the required output quality.

In Rodin Gen-2.5, fast mode focuses on efficiency and low cost. It supports batch generation, allowing users to produce multiple candidates at once and only pay when downloading. This is especially useful for rapid concept exploration, game asset prototyping, and real-time use cases, such as embedding 3D generation into games — for example, our collaboration with NetEase’s Eggy Party, which has reached over 500 million registered players globally and over 100 million monthly active users. 

High-detail mode is designed for most creative scenarios, especially complex or hero assets. It uses more compute to refine geometry and details — giving the model more “thinking time,” while the artist can take a coffee break while it works.

The update introduces 10M+ polygon generation for highly detailed surfaces and characters. What technical advancements made that possible, and where do you see this level of detail becoming most useful in production?

We think of 10M+ polygon generation like a “RAW file” in photography — it preserves maximum geometric information before any compression.

In production, this gives artists more flexibility downstream. A high-density 3D model can be used as a source for baking high-quality normal maps, creating game-ready low-poly assets, preparing 3D printing models, supporting film and VFX asset processing, or continuing sculpting workflows in tools like ZBrush and Blender. 

It’s not meant to be used directly everywhere, but to serve as a high-fidelity source that can be adapted to different pipeline needs. 

Of course, all of this only works if you actually have a model capable of faithfully reconstructing that level of detail lol.

For game developers specifically, high-poly output is often only one stage of the workflow. How does Rodin approach topology, optimization, and downstream use cases for real-time engines like Unreal Engine or Unity?

For game developers, high-poly output is not always a necessary stage in the workflow — especially for simpler games.

That’s why one of the most exciting updates in Gen-2.5 is Smart Low-poly, which we highlighted at the end of our demo.

It uses a GPT-like autoregressive approach to reconstruct meshes face by face, producing artist-style triangle and quad geometry that works well for real-time use cases.

This feature is still in beta, and we’re actively improving its performance.

Even before that, Rodin already supports standard low-poly outputs (a few thousand polygons), fully compatible with Unity and Unreal, with plugins for direct use in-engine.

Our DCC Bridge also integrates with major tools like Blender, Maya, Godot, and Cinema 4Dso artists can bring AI-generated 3D workflow into the software they already use seamlessly. 

AI-generated textures often struggle with seams, blurred backsides, or inconsistent details in unseen areas. What is different about Rodin Gen-2.5’s “3D-native” texture generation approach compared to more traditional AI workflows?

Rodin Gen-1 was the first to bring truly native 3D geometry generation into a production product — a direction that has since become mainstream, with many competitors following this paradigm.

However, up until Gen-2.5, texture generation in most systems still relied on a 2D workaround: using image generation models to produce multi-view renders, which were then projected back onto the 3D mesh to form textures. This often leads to issues like multi-view inconsistency, seams, and projection artifacts in occluded regions. The main reason for this approach is that native 3D texture modeling is significantly harder and constrained by limited training data.

With Gen-2.5, we moved texture generation to a truly 3D-native paradigm, where colors and materials are generated directly on the surface of the 3D model itself. This enables full-angle consistency, better controllability, and significantly improved PBR quality.

We will be releasing a technical report on this breakthrough later this year.

Rodin includes both “Faithful” and “Creative” generation modes. How do you define the difference between those approaches, and what kinds of artists or production scenarios benefit most from each?

During Rodin Gen-2, we found that many users input AI-generated images, which often contain subtle perspective inconsistencies that are hard to notice but can strongly affect 3D reconstruction.

That’s why we introduced “Creative” mode, which is more robust to imperfect inputs. “Faithful” mode is better suited for real-world, physically consistent images.

In most AI-driven workflows today, we generally recommend Creative as the default.

Batch generation appears to be a major focus in Gen-2.5. How do you see parallel generation changing the way concept artists, technical artists, and environment teams explore variations or build asset libraries?

Randomness is a key strength of generative AI, but it also forces artists to spend time iterating and selecting results. Parallel generation helps solve this by improving efficiency.

In Rodin Gen-2.5, users can generate up to 10 variations in a single run, each with subtle differences — instead of running 10 separate iterations. This is especially useful for concept artists, environment artists, game teams, and technical artists who need to quickly compare shapes, silhouettes, props, or asset variations. 

Looking forward, we aim to learn user preferences over time so the system can better adapt to each creator’s style.

One of the more interesting additions is “Manual BANG to Parts,” which allows creators to intentionally separate generated models into controllable components. Why is part-level separation such an important step toward production-ready AI 3D?

In most 3D workflows, assets are expected to be properly split into parts rather than kept as a single mesh. For example, game characters often need separate armor, clothing, weapons, accessories, or hair components, while 3D printing workflows often require models to be divided for assembly, material choice, or manufacturing constraints. 

In Gen-2, automatic BANG-based part decomposition was widely used, but users also wanted more control over how models are segmented, which raised the bar for controllability.

Manual BANG to Parts gives artists more intentional control over part separation. Combined with Part Refine, it allows users to clean up or enhance specific local areas without regenerating the entire model. 

Since 3D spans many industries, enabling manual, controllable part separation makes AI-generated assets much more suitable for real production pipelines, including 3D printing, animation, game development, and technical art workflows. 

There’s currently a lot of discussion around AI replacing artists, but your positioning seems more focused on accelerating workflows and improving iteration speed. How do you personally view the relationship between AI tools and professional artists?

What makes our team different from many other AI companies is that all of our co-founders are both generative AI researchers and 3D artists. We’ve all worked on film and game projects, and we are all Blender users ourselves — so the relationship between AI and artists is something we deeply care about.

As you mentioned, AI should accelerate workflows and improve efficiency, not replace artists. It should be a tool for artists, which is exactly why we place so much emphasis on controllability.

We believe AI companies should actively support the growth of the 3D industry and respect artists’ work. That’s also why we are a Golden Sponsor of Blender, and why all of our training data is properly licensed — including datasets purchased at scale from platforms like Shutterstock.

For professional artists, the real value of AI 3D generation is not removing creative judgment, but reducing repetitive work, speeding up early exploration, and giving artists more time for direction, refinement, and final quality control. 

Only AI built within this kind of healthy creative ecosystem can truly become a usable, production-grade tool that artists are willing to adopt.

What can AI 3D generation genuinely do well today, and where do you still think artists and studios should remain cautious or skeptical?

Today, AI 3D generation is already very strong at fast ideation, producing initial assets, exploring variations, and accelerating early-stage workflows. It can quickly turn concepts, sketches, text prompts, or reference images into usable 3D starting points and significantly reduce the cost of iteration. 

However, artists and studios should still be cautious when it comes to final production quality, especially in complex cases requiring strict topology, animation-ready edge flow, or highly specific art direction. AI outputs can still be inconsistent, and they often need human refinement to meet production standards.

In short, AI is excellent for speeding up the creative process, but it works best as a collaborator with artists, not as a substitute for professional artistic judgment or pipeline-level control. 

Looking ahead, do you think the future of AI 3D is primarily about improving generation quality, or is the bigger opportunity deeper integration into existing production pipelines and creative workflows?

I think generation quality will continue to improve, but that alone is not the main bottleneck anymore.

The bigger opportunity is deeper integration into real production pipelines and creative workflows — making AI actually usable inside how studios already work, rather than sitting as a separate “generation tool.”

That includes better controllability, compatibility with engines and DCC tools, and the ability to support iterative, artist-driven processes at scale.

For Hyper3D Rodin, the future of AI 3D is not only about creating a better model from a prompt. It is about building a controllable AI-to-3D workflow that can connect with game development, VFX, real-time rendering, 3D printing, and enterprise production pipelines. 

Finally, for 80 Level readers working in games, VFX, real-time rendering, or technical art, what workflow would you most recommend experimenting with first when trying Rodin Gen-2.5?

Rodin Gen-2.5 is designed to bring more control into the full AI-to-3D production pipeline, from generation speed and geometry complexity to texture quality, local detail, part-level editing, topology, and downstream optimization. 

For teams working on simpler game assets or fast concept iteration, I’d recommend starting with the Low or Extreme-Low thinking-effort modes, combined with 5x or 10x batch generation. These modes can generate multiple options in just a few seconds, around 4 seconds for Extreme-Low and 9 seconds for Low, making it much easier to explore variations quickly. From there, users can apply Smart Low-Poly to get a cleaner, more production-friendly mesh for real-time use in engines like Unity or Unreal Engine.

For more complex characters, creatures, or hero assets, the Extreme-High mode performs very well. It can capture much richer geometry and surface detail, especially when paired with detailed or micro presets. This is useful for high-detail sculpting references, VFX assets, 3D printing, and high-fidelity source models that can later be optimized or baked into lower-poly versions. 

Beyond the web platform, we also offer API access, group accounts, dedicated clusters, and private deployment options, so studios can scale Rodin Gen-2.5 into internal workflows depending on their production needs. 

Ready to give it a go? 80 Level readers can try Rodin Gen-2.5 free for 14 days using code 80lvRodinGen25.

QX Zhang, CTO of Hyper3D