Back to Gap Map Model components / Image, video, 3D & music generation

InstantMesh

Tencent
restricted / Overall score: 2.6

InstantMesh is Tencent ARC Lab's image-to-3D model. A multi-view diffusion model first draws several views of the object in a single image, and a sparse-view reconstruction model then turns them into a textured 3D mesh, the whole run taking about ten seconds.

Released April 2024; the repository was last updated in January 2025 and no later InstantMesh release exists. The multi-view stage is a fine-tuned Zero123++ v1.2, whose base pipeline the inference script downloads from SUDO AI's Hugging Face account.

Openness

2 medium confidence
2.0
weights
open (the four reconstruction checkpoints and the fine-tuned Zero123++ UNet download from TencentARC/InstantMesh without a gate)
data
described (trained on Objaverse renders; the README says the training set is not provided because of its size)
code
open (training code for the reconstruction models and for the Zero123++ fine-tuning is published)
license
Apache-2.0 (the InstantMesh code and weights, per the LICENSE and the model card)+CC-BY-NC-4.0 (Zero123++ v1.2, which run.py downloads by default for the multi-view stage and which the InstantMesh UNet fine-tunes)

InstantMesh's own code and weights are Apache 2.0, but its first stage runs Zero123++ v1.2, whose weights are licensed for non-commercial use only, and the inference script downloads them by default. Training code is published; the Objaverse renders it trained on are not.

Adoption

2 high confidence
2.0

Adoption is measured as Hugging Face downloads of TencentARC/InstantMesh, the only first-party copy of the weights.

Capability

3 medium confidence
3.0

InstantMesh places in the upper half of the image-to-3D arena for rendered quality, just behind Hunyuan3D-2 and 2.1, and in the lower half of the board that rates mesh topology.

Verified 2026-11-12