Feed-forward speed
Paper and Space target ~10-second mesh creation versus long SDS-style optimization loops.
TencentARC · Open image-to-3D
InstantMesh
Feed-forward multi-view + sparse-view LRM meshes in seconds (TencentARC)
InstantMesh is a feed-forward framework for efficient 3D mesh generation from a single image. It combines multi-view diffusion with a sparse-view large reconstruction model (LRM-style) so diverse assets can be created within about 10 seconds—see arXiv:2404.07191 and the official Space description.
What is InstantMesh?
Unlike optimization-based image-to-3D that can take minutes, InstantMesh is designed as a feed-forward path: first synthesize multi-view images with a multi-view diffusion model (Zero123++ variant in the release), then reconstruct a mesh with a transformer-based sparse-view reconstructor.
To improve training efficiency and exploit depth/normal supervision, the authors integrate differentiable iso-surface extraction and optimize on mesh representations directly. They release multiple reconstruction model variants and weights on Hugging Face (TencentARC/InstantMesh).
The official Gradio Space advertises “Create a 3D model from an image in 10 seconds!”—matching the paper’s seconds-scale claim for the full pipeline.
When to use InstantMesh
Use InstantMesh when turnaround time and mesh usability matter more than maximum high-res generative fidelity—rapid props, character shells, and interactive demos.
How InstantMesh works
STEP 1
Single image in
Upload an object image (white-background friendly workflows are discussed in the repo).
STEP 2
Multi-view synthesis
A multi-view diffusion model invents additional viewpoints that stay 3D-consistent enough for reconstruction.
STEP 3
Sparse-view LRM mesh
Transformer reconstructor predicts a mesh from the sparse views in a single feed-forward pass.
Key capabilities
Multi-view then reconstruct
Off-the-shelf multi-view diffusion produces consistent views for a sparse-view LRM reconstructor.
Mesh-centric training
Differentiable iso-surface extraction enables direct mesh optimization with geometric supervision.
Open ARC release
Code, weights, and demo from TencentARC; multiple reconstructor size variants on the model card.
Technical snapshot
Figures below come from the linked papers, model cards, or project pages—not from unbenchmarked third-party marketing claims.
- Paper
- arXiv:2404.07191
- Pipeline
- Multi-view diffusion → sparse-view LRM mesh
- Latency (authors / Space)
- On the order of ~10 seconds
- Org
- TencentARC
- Weights
- huggingface.co/TencentARC/InstantMesh
- Demo
- spaces/TencentARC/InstantMesh
Who it is for
Rapid prototyping
Block out game props or VRChat-style shells quickly before retopo.
Interactive demos
Seconds-scale feedback for product pages and internal tools.
Teaching / onboarding
Classic open example of multi-view diffusion + LRM reconstruction pipelines.
Limitations & practical notes
- Multi-view diffusion can invent incorrect back sides; expect artifacts on rare poses or thin structures.
- Fidelity is typically below larger 2025–2026 generative systems (TRELLIS.2 / Hunyuan3D-2.1) on hard assets.
- Exact runtime depends on Space GPU class and queue.
Frequently asked questions
Is InstantMesh the same as Instant3D / LRM?+
It builds on LRM/Instant3D-style sparse-view reconstruction ideas but is a distinct TencentARC system that pairs multi-view diffusion with mesh-optimized reconstructors (arXiv:2404.07191).
Where can I get weights?+
Official model card and GitHub: TencentARC/InstantMesh on Hugging Face and GitHub.
Sources & references
Landing claims are grounded in the following primary materials. Always prefer the original paper or model card for citations.
Try InstantMesh in the browser
Jump into the playground and generate a 3D model from your own image—no install required.