Recently, the field of 3D generation AI has seen another breakthrough. Hyper3D (Deemos Technology) officially launched Rodin Gen-2.5, claiming to be the strongest 3D generation model to date. It not only achieves generating a million-face model in 4 seconds but also breaks through to become the world's first 10 million face-level 3D generation, with details reaching the level of pores and skin microstructures, marking a crucial step for AI3D from "viewable" to "usable."

image.png

Double Breakthroughs in Speed and Scale: Adaptive Thinking Mode is the Highlight

Rodin Gen-2.5 introduces an **Adaptive Thinking Effort** mechanism similar to LLMs, dynamically allocating computing resources based on task complexity, offering multiple generation modes:

  • Very Low Mode: 4 seconds to generate, suitable for quick production of simple assets and mass testing
  • Low Mode: 9 seconds to generate, suitable for minimalist style models and small hard-surface props
  • Middle Mode: 20 seconds to generate, balanced structure and detail
  • High Mode: 40 seconds to generate, producing high-quality assets with rich structural layers and smooth surfaces
  • Very High Mode: 80 seconds to generate, focusing on professional-grade assets with microscopic details

AIbase pointed out that this tiered design significantly improves actual production efficiency, allowing users to choose flexibly according to their needs, enabling both fast iteration of ideas and high-fidelity asset output.

Native 3D Textures: 360° No Blind Spots, Saying Goodbye to Blurry Texture Maps

Differing from the common 2D-up-to-3D approach in the industry, Rodin Gen-2.5 adheres to a native 3D generation architecture. Just one reference image is needed to directly generate a complete texture model with PBR materials:

  • Textures are generated in 3D space, achieving 360° full coverage without distortion or blurring on the back or bottom
  • Strictly aligned with geometric structures, subtle details like fabric texture and seams are naturally handled
  • Supports Faithful Mode (strictly matching the reference) and Creative Mode (automatically optimizing structure, such as correcting tires into perfect circles)
  • At the highest precision, it can switch between Micro (microscopic details at the pore level) and Clean (clean and smooth geometry, suitable for stylized or animation purposes)

In addition, the model supports generating up to 10 results in parallel, greatly facilitating batch exploration of different creative directions.

Domestic Team Focused on Native 3D: Award-Winning Paper from SIGGRAPH 2025

The **Deemos Technology** behind Rodin is a domestic team focused on 3D generation. Since 2016, they have been deeply engaged in this field. While the industry generally adopted the 2D-to-3D shortcut, they chose a more challenging but thorough native 3D path, effectively solving long-standing issues such as broken faces and messy topology.

This year, the team's paper won the SIGGRAPH 2025 Best Paper Award, with only Google and Meta winning among commercial companies, further proving its technical strength.

AIbase analysis suggests that with the release of Rodin Gen-2.5, AI3D generation is accelerating towards production readiness. It has great potential for application in game development, e-commerce, AR/VR, and 3D printing. In the future, the barrier to 3D content creation is expected to significantly decrease, providing professional creators and developers with more efficient productivity tools.