You need a photoreal coffee mug for a game scene. You can model it in an afternoon, or generate it in twenty seconds — but the generated one looks like plastic under every light. The gap between "a shape" and "a thing that behaves like an object" is exactly what Qunhe Technology (群核科技) is chasing with Lux3D.
The Hangzhou company, one of the so-called "six little dragons" of the city's AI scene, released Lux3D on August 27, 2026. It is a text-to-3D and image-to-3D model built around one claim: realistic material matters more than geometry for getting generated 3D into real production.
What Lux3D actually outputs
Lux3D takes either a text prompt or a single uploaded image and returns 3D assets in three formats:
- Mesh — the standard polygon mesh most pipelines expect.
- PBR materials — physically based rendering data carrying metalness, roughness, transparency and refraction gloss, not just a painted color.
- 3D Gaussians — a fast-preview representation good for quick look-and-feel checks before committing to a full mesh.
A "standard" version and a "fast" version cover the two ends of the workflow: careful production versus rapid iteration.
Why material is the bottleneck
Most 3D generators have gotten decent at outline and silhouette. The hard part is surface behavior. The same red falls differently on glass, metal, ceramic or plastic, and a model that ignores that looks fake the moment light moves.
Lux3D trains on both generated data and large volumes of reconstruction data pulled from the physical world, which the company says constrains hallucinations and keeps materials believable across angles and lighting. In internal evaluations reported by state-linked media, it scored near the top on PSNR (pixel reconstruction accuracy) and CMMD (material realism).
Speed and the fast mode
The fast version is the headline for production scale: a single model builds in as little as about 20 seconds. That is the number that decides whether a studio can generate assets in bulk rather than by hand.
At that speed, the use cases shift. Game studios can populate environments with variant props; e-commerce teams can turn product photos into rotatable 3D displays in minutes instead of days; cultural and toy designers can iterate shapes quickly before committing to manufacture.
Where Lux3D sits among Chinese 3D generators
Lux3D enters a field already populated by VAST's Tripo series, Deemos' Rodin line and ByteDance's Seed3D. Qunhe's differentiator is the material-first framing and the explicit "drop into your existing tool" integration, rather than a race to the most detailed single sculpture. For a company whose main product is Kujiale (酷家乐), a home-design platform, the logic is clear: 3D assets that behave like real materials are immediately useful for interiors, furniture and spatial design.
Harness mode and pipelines
Qunhe is explicit that Lux3D is meant to sit inside existing tools, not replace them. Through a "Harness mode," developers can call it from coding assistants such as Codex and from modeling tools such as Blender — a single sentence in an AI coding workflow can batch-generate 3D assets instead of hand-building them.
That positioning matters. The 3D generation field is crowded with impressive demos and short on assets that drop cleanly into Unreal, Unity or a factory CAD chain. A model that ships an API into the tools people already use is answering the "then what" question directly.
The world-model teaser
Lux3D is one tile in a larger mosaic. Qunhe is internally testing a next-generation world model built on last year's SpatialGen spatial-generation model and on Lux3D itself, aiming for interactive, editable 3D worlds usable in video, games, embodied-agent training and digital twins.
According to Li An, chief scientist at BrainNet (脑机网), realistic material — not just geometry — is what finally lets generated 3D drop into real production pipelines.
Honest limitations
- "20 seconds" is a best case. The fastest mode is the floor, not the typical time; complex or high-fidelity assets will take longer, and the company has not published a full timing table.
- Benchmark figures are vendor- and media-reported. PSNR and CMMD rankings come from Qunhe and state-linked outlets; independent, standardized comparisons against Tripo or Meshy are not yet public.
- Material realism is still bounded by training data. Reconstruction data helps, but unusual materials or physically tricky combinations may still drift from reality.
- The world model is a teaser. The SpatialGen-based world model is described as in internal testing; no public release date or spec was given.
What readers can do now
- Test the material claim directly: generate the same object in mesh and in 3D Gaussian, then check it under several lighting angles to see if metalness and roughness hold.
- Wire it into Blender via Harness mode: if you already script assets, try a one-line batch call to replace a hand-modeled prop set.
- Use the fast mode for variants, the standard mode for heroes: bulk-generate background props at speed, then spend time on the few hero assets that carry the scene.