Rodin Gen 2.5, the Complete Guide
Gen-2.5 is Hyper3D's shift from "one-click generator" to controllable AI-to-3D workflow system - 10M+ polygon RAW output, 3D-native texturing, five adaptive speed tiers, and a full set of part-level editing tools. Here's what every feature actually does, and where it fits in a production pipeline.
What Gen-2.5 actually changed
Rodin Gen-2.5 launched May 19, 2026 as Hyper3D's production-control release. Where earlier Rodin models optimized mainly for speed and one-shot output, Gen-2.5 is built around a different pitch: a controllable AI-to-3D workflow system meant to sit alongside artists at every stage of production, not replace a single step of it. That shows up in three concrete places - adjustable "thinking" compute per generation, geometry detailed enough to treat as a RAW asset rather than a final export, and a set of part-level editing tools (ControlNet, Partial Edit, BANG to Parts) that let you steer or fix a result instead of regenerating from scratch.
The headline numbers from launch: 1 million polygons generated in about 4 seconds, and a maximum RAW output north of 10 million polygons - described internally as the first 3D generative model to reach that scale. See the full Rodin Gen-2.5 product page for the official feature summary, or the plan-by-plan breakdown on the pricing page for which Gen-2.5 features are gated by tier.
Five speed tiers, modeled on how LLMs reason
Gen-2.5 borrows the "thinking effort" concept from reasoning LLMs: more compute time buys more geometric and textural detail. You pick the tier per generation, so quick concept passes and hero-asset finals can sit in the same project without switching models.
| Tier | Typical time | Best for |
|---|---|---|
| Extreme-Low | ~4 seconds | Rapid batch concepting, silhouette checks |
| Low | ~9 seconds | Fast iteration on a prompt or reference image |
| Medium | ~20 seconds | Balanced default for most working assets |
| High | ~40 seconds | Production-ready detail, improved further by the 0702 patch |
| Extreme-High | ~80 seconds | Hero assets and maximum-detail RAW output |
The same Extreme-High effort concept also applies to texturing on its own axis - setting Texture Thinking Effort to Extreme-High is what unlocks the 12K Texture Generator beta, covered below.
10M+ polygon RAW output, then Smart Low-Poly
Hyper3D frames the 10M+ polygon RAW mesh the way a camera RAW file works in photography: it's the maximum-information version of the asset, not something you ship directly. From there, two paths bring it down to a usable range:
- Manual decimation / export controls - bring the RAW mesh down to whatever face count (roughly 1K-2M faces) and topology (Raw or Quad) a target engine or renderer needs.
- Smart Low-Poly (beta) - an autoregressive, face-by-face reconstruction that rebuilds the asset as artist-style low-poly geometry with smoother, more intentional surfaces than typical auto-decimation, aimed specifically at real-time and game-engine use. See the dedicated Smart Low-Poly guide for details, and note it's not available on the Free plan - covered in the free plan overview.
For most day-to-day assets, Medium or High effort at a few thousand to a few hundred thousand polygons covers real-time work; the full 10M+ RAW ceiling matters mainly for sculpting references, VFX hero assets, and high-fidelity 3D printing.
3D-native texturing, and how far the resolution goes
Earlier texture-generation approaches painted a model by projecting 2D images onto it from several camera angles, which tends to leave seams and inconsistent detail wherever those views overlap or miss a surface entirely. Gen-2.5 instead generates color and material directly on the 3D surface - Hyper3D calls this eliminating the "blind spot" problem - and it can preserve stylistic cues from a reference image, including toon-shaded looks, rather than flattening everything toward photoreal.
Standard texture output
Included on Creator and Business plans as the default resolution.
Free for every plan
Unlocked in the "0702" patch - previously a paid-only resolution, now free across all tiers.
Texture Generator, beta
Set Texture Thinking Effort to Extreme-High. Limited to Business-plan subscribers.
ControlNet, Partial Edit, and part-level tools
This is the part of Gen-2.5 that separates it from a pure text-to-3D generator: a set of tools for steering a result before it's final, or fixing one part of it after.
3D ControlNet
Three spatial guidance methods constrain shape and proportion before generation: a Bounding Box ControlNet for overall silhouette and scale, a Voxel ControlNet for coarse volumetric structure, and a Point Cloud ControlNet for guiding generation from existing 3D scan or reference data.
Partial Edit
Upload an existing mesh and modify a selected region with a natural-language prompt, without rebuilding the whole asset. See the standalone Partial Edit guide for worked examples.
BANG to Parts & Part Refine
BANG-based segmentation splits a generated asset into discrete components automatically, or manually via click-to-separate ("Manual BANG to Parts") when you need a specific breakdown - useful for rigging, per-part texturing, or isolating a piece to fix. Part Refine then lets you improve a single part's detail locally instead of regenerating the full model.
Free Retry & generation modes
Regenerating or redoing a result before you download it doesn't cost credits - see the pricing and credits breakdown for how that pay-by-result system works. Gen-2.5 also exposes two generation modes: Faithful, tuned for physically consistent results from clean real-world reference images, and Creative, tuned to handle imperfect or stylized inputs more robustly.
Up to 10 results per run
Gen-2.5 can generate up to 10 variations in a single run, which matters most at Extreme-Low or Low effort for fast concept exploration - generate a batch, pick the direction that works, then push just that one result through a higher effort tier for final detail. Input-wise, Rodin accepts up to 5 reference images per generation; see multi-image reference generation for how multiple angles improve accuracy over a single image.
Formats and where Gen-2.5 plugs in
Exports cover the formats most pipelines expect: STL, FBX, OBJ, GLB, GLTF and USDZ - full details on the export formats page, including which formats carry texture and rig data. On the integration side, Gen-2.5 ships native plugins for Blender, Unity, Unreal Engine, Godot, Maya, 3ds Max, Cinema 4D and Omniverse, so a generated asset can land directly inside an existing DCC or engine project rather than round-tripping through manual file imports - see the Blender integration guide for a walkthrough of that workflow.
What changed since Gen-2
| Capability | Gen-2 | Gen-2.5 |
|---|---|---|
| Speed control | Fixed generation time | 5 Adaptive Thinking Effort tiers, ~4s-~80s |
| Max polygon detail | Standard real-time range | 10M+ polygon RAW output |
| Texturing method | Multi-view 2D projection | 3D-native, no projection blind spots |
| Local editing | 3D inpainting via box-select + prompt | Partial Edit + BANG to Parts + Part Refine |
| Shape guidance | Limited | Bounding Box / Voxel / Point Cloud ControlNet |
| Batch output | Single result per run | Up to 10 results per run |
| Low-poly optimization | Manual decimation | Smart Low-Poly (beta), autoregressive |
Gen-2's 3D inpainting (box-select + text-prompt editing) shipped February 9, 2026 and became the foundation Partial Edit builds on - see the full release timeline for every dated update.
Where teams are putting Gen-2.5 to work
Game assets
Ideation, prototyping and low-poly production props
Characters & creatures
Reference-driven generation with ControlNet guidance
VFX hero assets
Extreme-High effort plus 10M+ polygon RAW detail
3D printing
High-fidelity RAW export for print-ready models
Real-time & XR
Smart Low-Poly output tuned for engine performance
Environment libraries
Batch generation for props and set-dressing at scale
Sculpt references
High-fidelity RAW meshes as a ZBrush starting point
Enterprise pipelines
API, group accounts, dedicated clusters, private deployment
See how this compares to hand-built and other AI pipelines in the game development workflow guide.
The "0702" patch
Gen-2.5's most recent update, rolled out in the weeks after launch, made three concrete changes:
- Better High-tier geometry - noticeable quality gains at the ~40-second High effort setting specifically, without moving to Extreme-High.
- 8K textures free for everyone - previously paid-only, now included on every plan including Free.
- New "Bloom" emissive texture mode - adds glowing/emissive material support to the texture generator.
For the full dated history - Gen-2's 3D inpainting, the Gen-2.5 launch itself, the 12K Texture Generator beta, and this patch - see the Rodin updates and release timeline.
Gen-2.5 walkthroughs
Tutorials and independent tests covering the features above, end to end.
Every Rodin guide on rodin3ds.com
Gen-2.5 features
Pricing & getting started
Gen-2.5: frequently asked questions
What is Rodin Gen-2.5?+
Rodin Gen-2.5 is Hyper3D's production-focused 3D generation model, launched May 19, 2026. It generates up to 10M+ polygon RAW meshes, uses five Adaptive Thinking Effort tiers (4-80 seconds), applies 3D-native texturing, and adds controllability tools like ControlNet, Partial Edit and BANG to Parts segmentation.
How fast is Rodin Gen-2.5?+
It depends on the effort tier: Extreme-Low generates in about 4 seconds, Low in about 9 seconds, Medium in about 20 seconds, High in about 40 seconds, and Extreme-High, the maximum-detail tier, takes about 80 seconds. Hyper3D has also demonstrated 1 million polygons generated in around 4 seconds.
How many polygons can Rodin Gen-2.5 generate?+
Gen-2.5 can output RAW meshes up to roughly 10 million polygons, intended as an uncompressed maximum-detail baseline similar to a photography RAW file. Smart Low-Poly (beta) then optimizes that detail down into clean, game-ready low-poly geometry.
What is 3D-native texture generation?+
Instead of projecting 2D images onto a mesh from multiple viewpoints, Gen-2.5 paints color and material directly on the 3D surface. This avoids the seams, occlusion blind spots and inconsistencies that multi-view projection texturing is prone to, and can preserve stylistic cues like toon shading from a reference image.
What is 12K Texture Generator and who can use it?+
12K Texture Generator is a beta feature that produces ultra-high-resolution textures by setting Texture Thinking Effort to Extreme-High. As of the most recent research, it's limited to Business-plan subscribers, while 8K textures became free for all plans in the mid-2026 "0702" update.
What is BANG to Parts?+
BANG to Parts is Gen-2.5's segmentation tool for breaking a generated asset into discrete components. It runs automatically by default, or manually via click-to-separate, so individual parts can be exported, retextured or refined with Part Refine without regenerating the whole model.