Rodin Gen-2
Last reviewed
May 16, 2026
Sources
20 citations
Review status
Source-backed
Revision
v1 ยท 2,768 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 16, 2026
Sources
20 citations
Review status
Source-backed
Revision
v1 ยท 2,768 words
Add missing citations, update stale details, or suggest a clearer explanation.
Rodin Gen-2 is a generative artificial intelligence model for producing three dimensional assets from text prompts or reference images, developed by the Shanghai based company Deemos and delivered through the company's Hyper3D creation platform at hyper3d.ai. The model was first opened for trials on 5 August 2025 and received its formal commercial launch on 1 October 2025, following a debut at SIGGRAPH 2025 where Deemos won the Best Paper Award for related research. It scales the underlying architecture to roughly 10 billion parameters and introduces a recursive part based generation method built on the team's BANG research, which lets the system produce complex objects as a set of clean, separable components rather than as a single fused mesh.
Rodin Gen-2 succeeds Rodin Gen-1, which Deemos opened for public trials on 30 May 2024 with a 1.5 billion parameter model, and the intermediate Rodin 1.5 release of 29 April 2025. A revised checkpoint dated 20 November 2025 added native 3D texture generation aimed at removing visible UV seams. The product is positioned for game development, visual effects, e commerce visualisation, and consumer 3D printing, and competes with Tencent Hunyuan 3D, Tripo P1 from Tripo AI, and Meshy 6 in the wider 3D generation market.
Deemos was founded in 2020 by Wu Di and Zhang Qixuan, two students from ShanghaiTech University, and is headquartered in the China Shanghai Pilot Free Trade Zone. The company first attracted attention with WAND, a GAN based sketch to image consumer app released in 2020 that reached the top of the Apple App Store Graphics and Design category in mainland China and accumulated roughly 1.6 million users in two weeks. Between 2021 and 2023 Deemos pivoted to high fidelity 3D scanning, building a light stage style capture system called the Plenoptic Stage that recorded around 2,000 actors with sub micron geometric and material precision. Observing that paying customers preferred AI generated assets over scanned ones, the team moved into generative 3D from 2023 onwards.
The company runs the consumer and enterprise platform Hyper3D, which currently hosts three products: the ChatAvatar character generator, the Rodin general 3D asset generator, and the OmniCraft 3D tool suite. Deemos has been backed by Sequoia China Seed Fund, MiraclePlus, BlueRun Ventures, and DragonBall Capital across its early rounds, and the company secured a Series A round co led by Meituan DragonBall and ByteDance in 2025.
The Rodin product line started with the public trial of Rodin Gen-1 on 30 May 2024. That first version was built on what Deemos called a 3D native large model with 1.5 billion parameters, and produced quad faced meshes with physically based rendering materials in around 40 seconds per asset. Rodin Gen-1 was the first commercial generative 3D system to advertise direct usability for 3D printing, and Deemos reported reaching an annual recurring revenue of roughly one million US dollars within 45 days of launch.
Rodin 1.5 followed on 29 April 2025. It scaled the underlying model to roughly 4 billion parameters and tightened the topology of the output meshes. The 1.5 release also expanded the Hyper3D web platform with team workspaces, asset libraries, and an early image to 3D mode aimed at concept artists. Rodin 1.5 was the model that prepared the company's third SIGGRAPH appearance and provided the production baseline that Gen-2 was benchmarked against.
Deemos demonstrated Rodin Gen-2 live at SIGGRAPH 2025 in Vancouver in August 2025 and made early access available to existing Hyper3D subscribers from 5 August 2025. The full commercial launch followed on 1 October 2025, with a press release positioning Rodin Gen-2 as a 10 billion parameter native 3D foundation model. The 20 November 2025 update, sometimes referred to inside the platform as build 1120, introduced a native 3D texture pipeline that the company says eliminates the seam artefacts associated with the older multi view to UV baking approach.
The table below summarises the capabilities disclosed in Deemos materials and independent reviews of the model.
| Capability | Status in Gen-2 | Source |
|---|---|---|
| Image to 3D generation | Supported, single reference image | Deemos press release |
| Text to 3D generation | Supported, natural language prompts | Hyper3D platform |
| Recursive part based generation | Headline feature, derived from BANG | Deemos press release |
| Quad based mesh topology | Supported, advertised as a 4x quality lift over 1.5 | Deemos press release |
| Baked normal maps | Supported, for low polygon proxies | Deemos press release |
| HD PBR texture maps | Supported in beta, albedo, roughness, metallic | befores and afters coverage |
| Native 3D texture generation | Added in 20 November 2025 update | Deemos changelog |
| 3D ControlNet inputs | Bounding box, voxel, and point cloud | Deemos press release |
| Generative Exploded Dynamics view | Supported, derived from BANG | SIGGRAPH paper |
| Natural language model editing | Supported via Rodin Gen-2 Edit, uploaded mesh + prompt | AWN interview with CTO |
| Approximate generation time | One to two minutes per asset | gaga.art review |
| Export formats | FBX, GLB, OBJ | Hyper3D documentation |
| Engine compatibility | Unity, Unreal Engine, generic DCC pipelines | Deemos press release |
Deemos does not publish a public benchmark suite for Gen-2, so the headline four times improvement in mesh quality versus Rodin 1.5 comes from internal evaluation. Independent reviewers at gaga.art and PixelDojo report that the topology is noticeably cleaner than competing systems on hard surface objects, while highly detailed organic forms can still need manual refinement and exact proportions in the output should be treated as approximate rather than measured.
The Rodin Gen-2 backbone scales to roughly 10 billion parameters, an order of magnitude above the 1 billion parameter Rodin Gen-1 model and noticeably larger than the 4 billion parameter Rodin 1.5. The technical lineage runs through three peer reviewed papers from the Deemos and ShanghaiTech team. The first is CLAY, a controllable large scale generative model for 3D assets that received the Best Paper Honorable Mention at SIGGRAPH 2024 and introduced the team's 3D native diffusion pipeline based on a 3DShape2VecSet style representation. The second is CAST, a component aligned 3D scene reconstruction method from a single RGB image, which won the Best Paper Award at SIGGRAPH 2025 on 14 August 2025 after three prior nominations. The third is BANG, the technique that gives Rodin Gen-2 its part level behaviour.
BANG, whose full paper title is "BANG: Dividing 3D Assets via Generative Exploded Dynamics," was authored by Longwen Zhang, Qixuan Zhang, Haoran Jiang, Yinuo Bai, Wei Yang, Lan Xu, and Jingyi Yu, with affiliations at ShanghaiTech University, Deemos Technology, and Huazhong University of Science and Technology. The paper was published in ACM Transactions on Graphics and selected as one of the Top 10 Technical Papers Fast Forward presentations at SIGGRAPH 2025. SIGGRAPH attendees later voted it into the conference Top 10 Presentations list overall.
BANG treats 3D asset creation as a process of progressive decomposition. Given a target object, the model produces a tree of parts: legs, seat, backrest, and arms for a chair, for example, with each component generated as a separate mesh and the relative spatial arrangement preserved. The Generative Exploded Dynamics view in the Hyper3D interface visualises this decomposition by animating the parts outward from the assembled position, letting a user inspect, edit, or swap individual components. Inside Rodin Gen-2 the same machinery underpins the recursive part based generation feature, where the model can subdivide a complex object into coherent components and continue refining each component independently rather than producing a single fused mesh.
Rodin Gen-2 accepts three spatial conditioning inputs in addition to text or image prompts. Bounding box control fixes the gross dimensions of an asset, voxel control supplies a coarse occupancy grid that the diffusion process must respect, and point cloud control provides a sparse surface guide. Deemos positions this as a 3D analogue of ControlNet in the 2D image generation world. The platform exposes the three inputs through the same web interface used for plain text and image generation.
The model emits quad biased meshes with baked normal maps, which lets a high polygon shape display its surface detail on a low polygon proxy suitable for real time rendering. Texture output includes albedo, metallic, and roughness maps, with the 20 November 2025 update replacing the older multi view to UV baking step with a native 3D texture generator that paints the surface directly in three dimensional space. Deemos says this removes the dead angle seams that appear when a UV chart is laid out poorly, although this has not yet been independently benchmarked.
Hyper3D is the consumer and enterprise web application that hosts Rodin Gen-2, ChatAvatar, and OmniCraft. It is accessed at hyper3d.ai and is operated directly by Deemos. The Rodin generator inside Hyper3D supports text prompts, single reference images, and the three ControlNet inputs, and exports to FBX, GLB, and OBJ files with optional PBR texture sets. The platform also hosts a community gallery of generated assets and a credit ledger that tracks usage against subscription quotas.
The Rodin Gen-2 Edit feature, introduced after the October 2025 launch and discussed in an Animation World Network interview with CTO Zhang Qixuan, lets users upload an existing 3D model and modify it using natural language prompts. According to Zhang, the goal is to move from one shot generation toward iterative editing workflows that look closer to how human artists actually work. Hyper3D also supports an API path so that the same generation and editing primitives can be invoked from third party tools, and Rodin Gen-2 has been integrated into hosted inference platforms including WaveSpeedAI.
Deemos prices Rodin Gen-2 access on a tiered subscription model billed in US dollars per month, with all tiers granting full commercial usage of generated assets. The free tier provides a limited trial, while paid tiers add monthly credit allocations and access to advanced features. The table below summarises the disclosed plans as documented in independent reviews of the Hyper3D platform.
| Plan | Price per month | Monthly credits | Notes |
|---|---|---|---|
| Free | 0 USD | Trial allocation | Watermark free, commercial use permitted |
| Education | 15 USD | 30 | Discounted plan for students and educators |
| Creator | 30 USD | 30 | Standard individual plan |
| Business | 120 USD | 208 | Higher quotas and team features |
| Enterprise | Custom | Custom | Direct contract with Deemos sales |
New Hyper3D accounts received a 14 day Creator Plan trial at launch under the promotional code bnaRodinGen2, distributed through the befores and afters coverage of the Gen-2 announcement. The earlier Rodin Gen-1 product had used a generation is free, download is paid pricing model, where users only spent credits when they accepted a result and downloaded the file. Some elements of that model persist in the free tier of the Gen-2 era platform.
The table below positions Rodin Gen-2 against three commonly cited rivals in the generative 3D space as of May 2026. Figures come from each vendor's public documentation and from independent reviews where cited; capabilities not confirmed by primary sources are left blank.
| System | Vendor | Open weights | Parameters | PBR textures | Headline strength |
|---|---|---|---|---|---|
| Rodin Gen-2 | Deemos | No | About 10 billion | Yes, HD with native 3D texture | Part based generation, quad topology |
| Hunyuan 3D 3.0 | Tencent | Yes, community licence | Not disclosed | Yes, 4K | Hard surface geometry, free open weights |
| Tripo P1 | Tripo AI | No | Not disclosed | Yes | Organic shapes, generation speed |
| Meshy 6 | Meshy | No | Not disclosed | Yes | Stylised characters, mature web product |
Reviewers at the 3D AI Studio benchmark roundup and the Vset3D channel converge on a similar pattern. Hunyuan 3D is favoured for hard surface modelling and for the price of admission since the weights are free to download within the licensed territory. Tripo is praised for organic creatures and stylised content and for fast iteration. Meshy is rated highly for product polish, documentation, and security certifications including SOC 2 and ISO 27001. Rodin Gen-2 is positioned as the realism and structural quality leader for higher budget productions, and it is the only one of the four to advertise a true part level decomposition mode for downstream editing. Independent reviewers also note that Rodin's pricing is at the premium end of the market, especially for studios that need high credit budgets.
Industry coverage of Rodin Gen-2 has been broadly positive. befores and afters described the launch as a groundbreaking step for intuitive 3D creation, the Animation World Network ran two pieces on the model including a follow up interview with the Hyper3D CTO, and PixelDojo positioned the system as the new high water mark for professional generative 3D. Coverage repeatedly highlights three points: the 10 billion parameter scale of the model, the four times improvement in geometric mesh quality claimed against Rodin 1.5, and the part level generation behaviour inherited from the BANG paper.
Academic reception has tracked the SIGGRAPH 2025 publication cycle. The CAST paper won the conference's Best Paper Award on 14 August 2025, BANG was named in the Top 10 Technical Papers Fast Forward list and later voted into the SIGGRAPH attendee Top 10 Presentations, and the team's earlier CLAY paper continues to be cited in subsequent 3D generative AI work. Deemos used the conference appearance both for academic outreach and to demonstrate Rodin Gen-2 live at its booth, a combination that the company has used at SIGGRAPH every year since 2023.
The most common criticisms in independent reviews target three areas. The first is pricing, since the credit based subscription model can run up costs quickly for studios that produce hundreds of assets per month. The second is detail handling on highly organic forms, where reviewers note that hair, fabric folds, and very fine surface detail can still require manual cleanup. The third is the closed weights model, which means that users cannot run Rodin Gen-2 locally and must rely on the Hyper3D cloud or partner inference platforms. The competitive comparison with Hunyuan 3D, where Tencent ships open weights under a community licence, recurs in many of these reviews.