If you tried early AI 3D generation (the image-to-3D generation), you were probably driven away by misaligned multi-view geometry, messy topology, and stretched textures. Deemos Technologies' newly released Hyper3D Rodin Gen-2.5 aims to end that — it brings the language-model paradigm of "think first, then generate" into 3D, achieving the world's first 10M-polygon-class generation, 1M polygons in 4 seconds, and native 12K textures, while in its first month online, subscribers and ARR both grew more than 400% month over month. This article breaks down its core capabilities, controllability design, and real use cases, to help modelers, product designers, and embodied-AI developers judge whether it's worth paying for.

What Is Deemos Hyper3D

Hyper3D is the native-3D generative model family from Deemos Technologies (founded in 2020, with a team from ShanghaiTech University). Its underlying framework, CLAY, was among the first in the industry to abandon the "lift 2D into 3D" route in favor of training on native 3D data, and it earned a SIGGRAPH best paper nomination. The latest generation, Rodin Gen-2.5, introduces LLM-like Test-time Scaling into 3D generation, letting the model "think before it generates," with five thinking-depth levels ranging from quick drafts to high-precision assets.

In Rodin Gen-2.5's first month online, subscribers and ARR both grew more than 400% month over month, and the consumer side began to overtake B2B.

One-line analogy: Rodin Gen-2.5 is to 3D what the GPT/o series is to code — not "a good-enough image," but production-ready assets.

Core Features

10M-Polygon-Class Generation + 1M Polygons in 4 Seconds

Gen-2.5 is the world's first 3D generation model at the 10M-polygon scale, and in Extreme-High mode it resolves details like skin microstructure, texture, and pores; in fast mode it generates a 1M-polygon model in 4 seconds. This effectively closes the gap that used to separate 3D assets meant for "getting the idea across" from those ready to go straight into film or game post-production.

The five thinking-depth levels correspond to generation times from 4 to 80 seconds; switch as needed.

Five Thinking-Depth Levels

Borrowing from LLM Test-time Scaling, users choose the model's thinking time and depth:

• Fast mode: 1M polygons in 4 seconds, for quick drafts and concept validation
• Extreme-High mode: the 10M-polygon class, for high-precision final deliverables
• Intermediate levels: trade off quality against speed per scenario

This design lets one model cover the whole path from inspiration to delivery, with no switching between tools.

Native 12K 3D Textures

Traditional texturing generates materials via projection, which commonly suffers from color banding, bleeding, and blurry logos. Deemos' companion native 3D texture model generates materials in parallel with the geometry stage, keeping logos and text crisp, and supports physically based PBR materials. Combined with the concurrently released 12K native textures, geometric precision and material fidelity can now surpass real-world scans.

Strong Controllability: 3D ControlNet + Partial Editing + Part Splitting

This is Deemos' most important differentiator. On the platform you can:

• Control the length, width, height, and shape of the generation with its in-house 3D ControlNet
• Make natural-language partial edits to platform-generated or third-party models ("make the ears a bit smaller")
• Auto-split 3D assets into parts, with splitting supported again on parts, for articulable assets

Part-splitting extends 3D assets from "static" to "articulable," which matters especially for games and embodied AI.

The Deemos team's core view: the second half of 3D generation has moved from chasing visual quality toward balancing controllability, efficiency, and quality. The seemingly tedious back-and-forth confirmation flow is precisely why paying customers choose them.

Real-World Experience and Positioning

Strengths

• Controllability crushes image-to-3D approaches: trained directly on native 3D data, avoiding multi-view alignment errors
• Confirm step by step, pay only when satisfied: a sign of confidence in the model's usability
• Deep enterprise bench: ByteDance, Unity, Figma, and Canva are all on the client list, and its B2B revenue exceeds that of all other companies in its niche combined
• Dense academic grounding: 30+ papers at SIGGRAPH and other top venues since 2020, with CAST winning SIGGRAPH 2025 Best Paper (in this field, the only award-winning commercial companies are Google, Meta, and Deemos)

Model Iteration Cadence

Four major iterations in the past two years — a fast pace for the 3D space:

1. Rodin Gen-1: the world's first native 3D large model
2. Rodin Gen-2 (September 2025): added part-splitting
3. Rodin Gen-2 Edit (January 2026): the first 3D editing model, making AI 3D editable for the first time
4. Rodin Gen-2.5 (June 2026): introduced the thinking mechanism, the 10M-polygon class, and 12K textures

Use Cases

Deemos' product positioning is unambiguous: serve the professionals who actually pay first. On the consumer side the profile is Pro-C (professional consumers); on the enterprise side it spans several high-value scenarios:

Embodied AI deserves a special mention. It places particular demands on 3D assets: physical feedback (weight, colliders) and interactive assets (how parts move once split). Deemos has already shipped physical feedback and published academic work on interactive assets. Within the world-model landscape, Deemos has chosen the vertical but practical route of "making simulatable assets."

How to Get Started

• Platform subscription: visit hyper3d.ai to sign up and subscribe; billing is by model and generation volume
• Enterprise API / on-prem deployment: for enterprise-scale batch calls and customization
• Payment model: confirm the generation step by step and pay only when satisfied

If you work on e-commerce 3D, game assets, or robot simulation environments, Rodin Gen-2.5 is one of the few 3D generation models currently capable of production-ready output, and it merits evaluation as a production tool.