🏆 CVPR 2026 Oral & Award Candidate Paper (Top 1.8%)

SuperFit fits compact assemblies of SuperFrusta and other primitives to 3D shapes, built on top of SySL. See Install below and BibTeX.

• May 2026: Added an optional custom CUDA kernel for VarAxisSF. On the recorded B200 benchmark, it is up to 8.5x faster end-to-end than the dynamic torch.compile path, and up to 38.8x faster for forward-only evaluation. Use it with --ablation 8; see notes/custom_vasf_cuda.md for details.

git clone https://github.com/BardOfCodes/superfit.git
cd superfit
conda env create -f env.yml
conda activate superfit

This installs PyTorch 2.9.1 (CUDA 12.8), all core Python dependencies, and superfit itself in editable mode.

cubvh provides GPU-accelerated BVH queries and is required by the fitting pipeline.

git clone https://github.com/ashawkey/cubvh.git
cd cubvh
python setup.py install
cd ..

Kaolin is used for FlexiCubes meshing. See the kaolin installation docs for full details.

git clone --recursive https://github.com/NVIDIAGameWorks/kaolin.git
cd kaolin
pip install -r tools/build_requirements.txt -r tools/viz_requirements.txt -r tools/requirements.txt
export IGNORE_TORCH_VER=1
python setup.py install
cd ..

Note: IGNORE_TORCH_VER=1 is needed because kaolin's version check may not yet list PyTorch 2.9.

Before running any scripts, create a local path configuration:

cp superfit_config.example.json superfit_config.json

Edit superfit_config.json for your machine:

{
  "data_base": "/path/to/your/data",
  "project_base": "/path/to/your/projects/project_sf",
  "outputs_base": "/path/to/your/outputs"
}

superfit_config.json is ignored by git. You can also point to another config file with SUPERFIT_CONFIG=/path/to/superfit_config.json. All dataset and artifact locations are derived from these values. See notes/dataset.md for details on expected data layout.

python scripts/mesh_to_assembly.py --input_path <path> --save_dir <save-dir> --fastmode --save_html --save_edit_html --save_mesh

This will convert a watertight input mesh into a compact assembly of SuperFrusta. Use different --ablation options to generate assemblies of cuboids/superquadrics/supergeons etc. Use --ablation 8 to enable the custom CUDA VarAxisSF path after building the extension. Note that --fastmode saves torch compile artifacts at AOT_ARTIFACT_DIR as specified in superfit/utils/constants.py.

If the input mesh contains textures, you can run fit_texture.py to add textures to the primitive assembly. This will add 2D spherical textures to each primitive. We also have testset_fit_textures.py for running this process across multiple inputs.

python scripts/fit_texture.py --input_path <path-to-assembly-pkl> --save_html --save_edit_html

Finally, you can render short videos from any saved primitive_assembly.pkl. For a compact demo, render the optimization replay, an exploded/colorized assembly, and a final spiral view, then concatenate them:

python scripts/visualize/generate_opt_video.py \
  --input_pkl <path-to-assembly-pkl> --mode opt_seq --save_dir <save-dir>

python scripts/visualize/generate_opt_video.py \
  --input_pkl <path-to-assembly-pkl> --mode explode_color_compact --save_dir <save-dir>

python scripts/visualize/generate_opt_video.py \
  --input_pkl <path-to-assembly-pkl> --mode spiral --save_dir <save-dir>

python scripts/visualize/generate_opt_video.py \
  --input_pkl <path-to-assembly-pkl> --mode combine --save_dir <save-dir> \
  --combine_segments opt_seq explode_color_compact spiral

Additional modes include color_reveal, explode_spiral, explode_pause_fit_back, and legacy_opt.

Note that we don't generate html shaders for SuperQuadric since we don't have analytical sphere-traceable SDF functions for them. Additionally, please change the configuration in superfit/utils/config.py and superfit/visualize/config.py (camera/video defaults) if needed.

We also release pre-computed primitive assemblies on Hugging Face:

The dataset contains SuperFit outputs for Toys4K and PartObjaverse, including primitive_assembly.pkl, fit configs, manifests, and evaluation summaries. The release is useful if you want to inspect, render, or evaluate existing assemblies without running the fitting pipeline yourself. See notes/dataset.md for the layout and licensing notes.

Known issue: primitive_assembly.pkl artifacts are Python pickle files. Only load PKL files from trusted sources, because loading a pickle can execute arbitrary code.

1. Fit primitives to Toys4K

python scripts/testset_fit_primitives.py --start_ind 0 --end_ind 500 --fastmode --save_dir <save-path>

This will generate primitive assemblies for our Toys4K evaluation subset. You can additionally use --dataset partobjaverse to generate fits for the PartObjaverse dataset. To run Toys4K fitting in parallel across GPUs, launch the release script from the repository root:

bash job_scripts/all_toys4k.sh

2. Run Evaluation

Once the primitives are generated, you can run:

python scripts/testset_eval.py --input_path <save-path> --save_per_instance_metrics --start_ind 0 --end_ind 500 --include_semantic

For the semantic metrics, we require faiss as well as PartField. Add --include_semantic to evaluate the semantic metrics.

We also provide a complementary web app to explore the primitive assemblies - the fitting process, the metrics, the generated shapes etc.

The inferred primitive assembly can also be used as spatial guidance to generate higher fidelity meshes by combining our method with SpaceControl, by E. Fedele et al.

Find instructions to install and run this at superfit_app.

Details regarding the dataset are provided in notes/dataset.md. For details regarding the primitives check out notes/primitives.md. We also made quite a few improvements over the results after our CVPR submission. These are listed in notes/post_submission.md.

The notebooks/ folder contains a few IPython notebooks which demo different aspects of our method such as (a) primitive design exploration in notebooks/primitive_designs.ipynb, (b) curvature exploration in notebooks/mesh_curvature.ipynb, and (c) morphological decomposition in notebooks/msd.ipynb.

@misc{ganeshan2026superfit,
  title         = {Residual Primitive Fitting of 3D Shapes with SuperFrusta},
  author        = {Aditya Ganeshan and Matheus Gadelha and Thibault Groueix and Zhiqin Chen and Siddhartha Chaudhuri and Vladimir G. Kim and Wang Yifan and Daniel Ritchie},
  year          = {2026},
  booktitle     = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month         = {June},
}

This project was developed during an internship at Adobe Research. I (Aditya) am grateful to all co-authors for their invaluable contributions: Matheus Gadelha, Thibault Groueix, Zhiqin Chen, Siddhartha Chaudhuri, Vladimir G. Kim, Wang Yifan, and Daniel Ritchie.

Our sphere-tracing primitives and shaders draw heavily on foundational work by Inigo Quilez, and the broader ShaderToy community — with special thanks to Paniq for creating the superprimitive and uberprimitive functions.

Finally, Anton Mikhailov, Daichi Ito, and Luc Chamerlat from Adobe provided valuable feedback around primitive design and artist needs.

For questions reach out at adityaganeshan@gmail.com.