Mirror Illusion Art (AutoMIA)

[CVPR 2026] This is the official implementation repository for the paper "Mirror Illusion Art" in CVPR 2026. AutoMIA is a PyTorch3D-based system for designing 3D mirror illusion objects from two target images. Given a pair of supervision images, the code jointly optimizes a binary voxelized shape and view-dependent colors, then converts the learned volume into a printable mesh.

Authors

Xiaopei Zhu, Zeyuan Li, Jun Zhu, Xiaolin Hu

Showcase

Overview

AutoMIA combines:

• Dual-view differentiable volume rendering with orthographic or perspective cameras
• Shape-color decoupled optimization for stable inverse design
• Projection-aligned component pruning to suppress floating artifacts
• Interior voxel preservation and mesh smoothing for fabrication-friendly outputs

Installation

We now provide a reproducible conda environment file in environment.yml.

Conda environment

conda env create -f environment.yml
conda activate automia

If you prefer a one-command setup on Linux, you can also use:

bash scripts/setup_env.sh
conda activate automia

Environment notes

• Recommended platform: Linux with CUDA 12.1
• Recommended Python version: 3.10
• environment.yml installs PyTorch 2.4.1, torchvision 0.19.1, and PyTorch3D 0.7.8
• CPU execution is possible for some steps, but training and mesh generation are intended for GPU use

You can verify that the environment is visible to the project with:

python run.py --list_gpus

Quick Start

run.py is the main file for training and conversion.

The released example supervision images should be placed in example/. The command below is the shortest end-to-end path for reproducing AutoMIA with the two example images:

python run.py \
  --train --convert \
  --supervision_image1 example/<view1_image> \
  --supervision_image2 example/<view2_image> \
  --volume_size 128 \
  --n_iter 800 \
  --lr 0.05 \
  --azim1 0 --azim2 45 \
  --elev1 0 --elev2 -45 \
  --output_dir results

Replace <view1_image> and <view2_image> with the actual filenames of the two example images stored in example/.

What run.py does

By default, run.py can train a voxel model and convert the latest checkpoint into a mesh. Common patterns are:

Train and convert

python run.py \
  --train --convert \
  --supervision_image1 example/<view1_image> \
  --supervision_image2 example/<view2_image>

Train only

python run.py --train \
  --supervision_image1 example/<view1_image> \
  --supervision_image2 example/<view2_image>

Convert the latest checkpoint only

python run.py --convert

Convert a specific checkpoint

python run.py --convert \
  --model_path results/<run_name>/models/<checkpoint>.pt \
  --threshold 0.5 \
  --voxel_size 0.1

Run from a JSON config

python run.py --config_json configs/v256_default.json --train --convert

Inputs

AutoMIA expects two supervision images:

• --supervision_image1: target appearance for view 1
• --supervision_image2: target appearance for view 2

The training code loads both silhouette and RGB supervision from these images. RGBA inputs are supported; otherwise the code derives a binary foreground mask from luminance thresholding.

Outputs

After a successful run, you should expect:

• Training artifacts in results/binary_voxel_<timestamp>/
• Intermediate renders and density slices in results/binary_voxel_<timestamp>/images/
• A saved checkpoint in results/binary_voxel_<timestamp>/models/
• A mesh export in meshes/<model_name>/<model_name>.obj
• A companion .ply mesh export for downstream visualization

Useful Arguments

Argument	Meaning
--volume_size	Voxel resolution, currently 128 or 256
--n_iter	Number of optimization iterations
--lr	Learning rate
--azim1, --azim2	Azimuths for the two supervision views
--elev1, --elev2	Elevations for the two supervision views
--pts_per_ray	Number of ray samples in the differentiable volume renderer
--render_width, --render_height	Explicit render resolution override
--no_orthographic	Switch from orthographic to perspective cameras
--shape_ratio	Fraction of iterations allocated to the shape-first stage
--threshold	Marching-cubes threshold used during mesh extraction
--smooth_iterations	Number of smoothing iterations after mesh extraction

Method at a Glance

The implementation is centered on four ingredients:

• PAC: projection-aligned component pruning
• PWA: position-weighted adaptive suppression for background artifacts
• IVP: interior voxel preservation to avoid internal fracture
• SCD: shape-color decoupled optimization

The high-level data flow is:

flowchart TD
    inputImages[TwoInputImages] --> runPy[run.py]
    runPy --> trainer[main_enhanced.py]
    trainer --> renderer[volume_illusion/renderer.py]
    trainer --> voxelModel[volume_illusion/model.py]
    trainer --> pac[volume_illusion/component_pruning.py]
    trainer --> checkpoint[Checkpoint]
    checkpoint --> meshExport[volume_to_mesh.py]
    meshExport --> finalMesh[OBJ And PLY Meshes]

Loading

Repository Layout

AutoMIA/
├── README.md
├── environment.yml
├── requirements.txt
├── run.py
├── main.py
├── main_enhanced.py
├── volume_to_mesh.py
├── configs/
│   ├── v256_default.json
│   └── projection_pruning_scoring.md
├── example/
│   └── README.md
├── scripts/
│   ├── setup_env.sh
│   ├── convert_rgb_to_rgba.py
│   └── README_RGB_TO_RGBA.md
└── volume_illusion/
    ├── model.py
    ├── renderer.py
    ├── visualization.py
    └── component_pruning.py

Citation

If you find this repository useful, please cite the AutoMIA CVPR 2026 paper.