Introducing Prism-2
Real meshes, real materials, in 38 seconds. Our second generation 3D foundation model is coming to early access.
Dimensa teamProduct and research
When we released Prism-1 as a research preview last winter, we made one promise: every output should be something you could actually use. Not a turntable video. Not a point cloud that looks great until you try to rig it. A mesh you can drop into an engine and ship.
Prism-1 mostly kept that promise for simple props. Prism-2 keeps it for characters, products, vehicles and full scenes. It powers early access, and we will invite the waitlist in waves.
What is new
Native mesh decoding
Most 3D generators predict a volume or a set of views, then run marching cubes to pull out a surface. That is where the lumpy edges, holes and melted details come from. Prism-2 predicts mesh structure directly: vertices, faces and part boundaries. The result is crisper silhouettes and far fewer artifacts on thin features like straps, antennae and fingers.
On our internal evaluation set, Prism-2 reduces geometric artifacts by 41% compared to Prism-1, and produces watertight meshes 97% of the time without any post processing.
Materials, not paint
Earlier models baked lighting straight into a single color texture. That looks fine in the thumbnail and terrible under a different light. Prism-2 predicts a full physically based material per part: albedo, normal, roughness, metallic and, where relevant, emissive, clearcoat and sheen.
Because materials are predicted per part, you can regenerate just the leather on a chair or just the glass on a lantern without touching the rest of the model.
Much faster
A new distilled sampler and a rewritten inference stack brought our median prompt to model time from 71 seconds down to 38 seconds, including texturing. Retopology and rigging run in parallel, so most assets are export ready in under a minute.
How we trained it
Prism-2 was trained exclusively on assets we license or commission. We work with independent artists and asset libraries who opt in, and contributing artists receive an ongoing share of revenue. We think this is the only sustainable way to build creative AI, and we will keep publishing how the program works as it grows.
By the numbers
| Metric | Prism-1 | Prism-2 |
|---|---|---|
| Median generation time | 71s | 38s |
| Watertight outputs | 81% | 97% |
| Max texture resolution | 2K | 4K |
| Material channels | 1 | up to 7 |
What comes next
Over the coming months we will roll out Scene Composer for full environments, text to motion for custom animations and private style fine tuning for studios. You can follow along on our public roadmap.
If you are not on the waitlist yet, now is a great time to join. Founding members get priority access, lifetime pricing and 1,000 bonus credits at launch. We cannot wait to see what you build.