TRELLIS
3D Available Full comparison ↗ microsoft/trellis · by Microsoft
· structured-latents-rectified-flow
Pricing — 1 offering(s)
About this offering: Microsoft's open-weights TRELLIS (MIT license, weights on Hugging Face), offered by several hosts. The prices shown are fal's: TRELLIS v1 at a flat $0.02 per generation and TRELLIS 2 at $0.25–$0.35 depending on output resolution (fal model pages, checked 2026-10-11). There's no first-party API: use a host such as fal or Replicate, or self-host.
Image-to-3D (TRELLIS v1)
- $0.02 / generation
Current
2026-09-02 → present
TRELLIS v1 costs a flat $0.02 per generation on fal (fal model page, checked 2026-10-11), one of the cheapest…
TRELLIS v1 costs a flat $0.02 per generation on fal (fal model page, checked 2026-10-11), one of the cheapest hosted image-to-3D options. It uses rectified-flow transformers (up to 2B parameters, trained on about 500K 3D objects) and can output radiance fields, 3D Gaussians or meshes (GLB export).
Image-to-3D (TRELLIS 2, 1024p)
- $0.30 / generation
Current
2026-09-02 → present
TRELLIS 2, the higher-fidelity successor, costs $0.25 at 512p, $0.30 at 1024p and $0.35 at 1536p output on…
TRELLIS 2, the higher-fidelity successor, costs $0.25 at 512p, $0.30 at 1024p and $0.35 at 1536p output on fal (fal model page, checked 2026-10-11). The 1024p rate is shown.
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how TRELLIS fits into a cost-aware routing setup
See how →Capability profile
How TRELLIS rates across core capability dimensions, with the task-level evidence behind each rating.
Image-to-3D primary; good silhouette/detail preservation reported. A text-to-3D path exists in the research release but the common hosted endpoints are image-conditioned.
Produces textured meshes; no separate PBR / re-texture controls on the hosted endpoints.
No quad / low-poly / polygon-budget controls exposed.
Static mesh output; no rig/animation.
Multiple output representations (mesh / Gaussian splat / radiance field) is the flexibility story; input is image-first on hosted endpoints.
Seconds-to-a-minute on hosted GPUs; no pinned latency figure.
Ratings are estimated — limited independent data is available for this model.
Operator guidance
Reach for TRELLIS when you want an open-weights model with the cheapest hosted rate (v1, $0.02/gen) or splat/NeRF output, or when you plan to self-host. Use TRELLIS 2 ($0.25-$0.35 by resolution on fal) when fidelity matters. For a higher geometry ceiling or a full asset pipeline, Hunyuan3D or the first-party APIs (Tripo / Meshy) are stronger.
Use cases
- Lowest-cost open-weights image-to-3D ($0.02/gen on fal for v1)
- Workflows that need Gaussian-splat or radiance-field output, not just meshes
- Self-hosted 3D generation (MIT weights)
Limitations
- Research model — no first-party API; access is via aggregators (fal / Replicate) or self-host
- No rigging, remesh, segmentation, or PBR-material controls on hosted endpoints
- Hosted endpoints are image-conditioned; text-to-3D is not a common hosted path