Model catalog
Live snapshot of every spec the Dream Engine HTTP gateway advertises via GET /v1/models. Auto-generated from web/data/catalog.json; re-run make update-docs after the catalog changes.
Engine version: 0.1.5
Default price: $0.0005 per frame
Catalog generated at: 2026-05-04T16:40:10+00:00
DreamDojo 2B · GR-1
dreamdojo-2b-gr1 — flagship · benchmarked
DreamDojo 2B · GR-1 is a 2-billion-parameter action-conditioned video world model designed for humanoid robotics simulation, visual model-predictive control, and large-scale rollout data augmentation. It pairs a chunked rectified-flow diffusion backbone with token-level action conditioning over the Fourier GR-1's 384-dimensional control space, delivering bit-stable 480×640 video rollouts at sub-3-second engine latency on Hopper hardware. It represents a top-tier open-source world model for end-to-end humanoid planning, on-policy evaluation, and production-grade training-data pipelines.
| Attribute | Value |
|---|---|
| Provider | NVIDIA |
| Category | Robotics |
| Resolution (H × W) | 480 × 640 |
| Action dim | 384 |
| Chunk size | 12 frames |
| Default steps | 35 |
| Runner | chunked |
| $ / frame | $0.0005 |
| License | NVIDIA Open Model License (cosmos_predict2) |
Benchmark
vs. DreamDojo base (stock cosmos_predict2) — see the model detail page for the full diagram.
| Metric | Value | Δ |
|---|---|---|
| Wall (s) | 8.79 | 3.42× faster |
| Engine wall (ms) | 2,616 | — |
| PSNR (dB) | 22.346 | +0.641 |
| SSIM | 0.8049 | +0.0254 |
| LPIPS | 0.2065 | -0.0205 |
| Cost / rollout (USD) | $0.0096 | — |
Specs not in this list aren't yet benchmarked + validated on Dream Engine. The model roadmap tracks upstream candidates we're integrating.