Sub-second reconstruction
Paper headline latency is ~0.5s for textured mesh reconstruction; model card states under one second.
Stability AI · Open image-to-3D
Stable Fast 3D
SF3D: sub-second textured mesh reconstruction with UV unwrap (Stability AI)
Stable Fast 3D (SF3D) is Stability AI’s open large reconstruction model that turns a single object image into a textured, UV-unwrapped mesh. The research paper “SF3D: Stable Fast 3D Mesh Reconstruction with UV-unwrapping and Illumination Disentanglement” (arXiv:2408.00653) targets high-quality results in about 0.5 seconds.
What is Stable Fast 3D?
SF3D builds on the TripoSR lineage but adds techniques aimed at cleaner meshes, UV unwrapping, illumination disentanglement (“delighting”), and material parameter prediction so assets integrate more easily into game engines.
The Hugging Face model card describes SF3D as a transformer image-to-3D model trained to create a 3D model from a single image in under one second, trained with data including Objaverse-scale resources (card tags reference allenai/objaverse).
Important: Stability AI publishes SF3D under a non-commercial research-oriented license on the model card—read the license before commercial redistribution. The official Space is stabilityai/stable-fast-3d.
When to use Stable Fast 3D
Use SF3D when you need a fast, UV-friendly mesh from one image for blocking and lookdev—and you accept research/non-commercial license constraints on the open weights.
How Stable Fast 3D works
STEP 1
Single image encoding
A transformer-style reconstructor maps one object crop into a 3D representation.
STEP 2
Mesh + UV + materials
Decode to a mesh with UV atlas and textures, aiming to separate baked lighting from albedo-like signals.
STEP 3
Export to DCC / engines
Download demo outputs and import into Blender, Unity, or Unreal depending on export format support.
Key capabilities
UV-unwrapped textures
Explicit optimization for meshes with UV unwrapping rather than only vertex colors.
Illumination disentanglement
SF3D delights lighting and predicts material parameters for easier engine integration (paper focus).
TripoSR heritage
GitHub README positions SF3D as improving on TripoSR while keeping feed-forward speed.
Technical snapshot
Figures below come from the linked papers, model cards, or project pages—not from unbenchmarked third-party marketing claims.
- Paper
- arXiv:2408.00653
- Also known as
- SF3D (Stable Fast 3D)
- Latency (paper / card)
- ~0.5s paper; under 1s on model card
- Outputs
- Textured UV-unwrapped mesh; material parameters
- License
- Stability AI Non-Commercial Research Community License (see card)
- Demo
- spaces/stabilityai/stable-fast-3d
Who it is for
Rapid product / prop capture
Quick meshes for catalogs, AR try-ons, and prototype stores.
Game blocking
Low-latency assets for greybox levels when perfect topology is not required yet.
Speed baseline
Compare latency/quality against InstantMesh, Hunyuan, and TRELLIS-class models.
Limitations & practical notes
- Open weights are not a free commercial license—review Stability AI’s terms before shipping products.
- Single-view reconstruction still hallucinates unseen sides; not a replacement for photogrammetry.
- Remote Space runtime can exceed paper latency under load.
Frequently asked questions
How fast is Stable Fast 3D?+
The SF3D paper reports about 0.5 seconds for textured mesh reconstruction; the Hugging Face model card describes generation in under one second. Hosted demos may be slower due to queues and hardware.
Can I use SF3D commercially?+
Not automatically. Check the Stability AI Non-Commercial Research Community License (and any Space terms) on the official model card before commercial use.
Sources & references
Landing claims are grounded in the following primary materials. Always prefer the original paper or model card for citations.
Try Stable Fast 3D in the browser
Jump into the playground and generate a 3D model from your own image—no install required.