# guochengqian/Magic123

[ICLR'24] Official PyTorch Implementation of Magic123: One Image to High-Quality 3D Object Generation Using Both 2D and 3D Diffusion Priors

Repository: https://github.com/guochengqian/Magic123
Canonical: https://ross.abutalabs.com/products/magic123
Homepage: https://guochengqian.github.io/project/magic123
Language: Jupyter Notebook
License: Apache-2.0
License Family: permissive
Topics: image-to-3d, 3d-generation, diffusion-models
Last push: 2026-07-06T02:45:31+00:00

## Health v2 (maintenance only)
Score: 90/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 91, release rhythm 92, longevity 83
- inputs: {"age_days": 1164, "days_push": 58, "days_rel": 58, "gap_med": 0, "n_releases_24m": 2}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1623, forks 99 (observed 2026-08-28T04:05:12.622616+00:00)

## What it is
Magic123 is the official PyTorch implementation of an ICLR 2024 paper that generates high-quality textured 3D meshes from a single unposed image using joint 2D and 3D diffusion priors. It uses a two-stage coarse-to-fine pipeline optimizing a NeRF then a differentiable mesh, with a tradeoff parameter controlling exploration vs exploitation of the priors.

## Use cases
- generate a 3D model from a single photo
- reconstruct a textured 3D mesh from one image
- create 3D assets from concept art
- convert 2D images into 3D objects for games
- research on image-to-3D generation with diffusion priors
- control creativity vs fidelity in single-image 3D reconstruction

## When to choose
- you need to turn a single unposed image into a textured 3D mesh
- you want a tunable tradeoff between imaginative (2D prior) and precise (3D prior) geometry
- you are reproducing or building on ICLR 2024 image-to-3D research
- you have a CUDA GPU and Ubuntu environment

## When to avoid
- you need fast real-time 3D generation - optimization takes significant GPU time
- you lack a CUDA-capable GPU
- you need production-grade 3D asset tooling rather than research code
- you only have multi-view images, where classical photogrammetry is more appropriate

## Facets
- artifact type: library
- maturity: stable
- function: machine-learning, deep-learning, image-processing, graphics, simulation
- domain: computer-vision, graphics, artificial-intelligence, deep-learning
- platform: python
- tags: image-to-3d, 3d-generation, diffusion-models, nerf, pytorch, research-code, iclr-2024, linux, gpu

## Member repositories
- guochengqian/Magic123 (main) score 90

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:12.622616+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T03:48:49.591160+00:00, confidence not recorded.
  - readme: https://github.com/guochengqian/Magic123 (fetched 2026-08-28T04:05:12.622616+00:00, sha 3eeff3b5bcfb)
  - homepage: https://guochengqian.github.io/project/magic123 (fetched 2026-08-29T11:21:37.656859+00:00, sha 701a9ea05dac)
  - registry_pypi: https://pypi.org/pypi/magic123/json (fetched 2026-08-29T11:21:37.666516+00:00, sha 4bbd72b2bda2)
- Data as of 2026-08-30T08:39:29.467469+00:00.