RADmesh: Remesh-Aware Mesh Deformation

The official implementation for our ECCV 2026 (Oral) paper.

project page | paper (pdf) | BibTeX

---

Installation

System requirements

Optimizations have been tested to run on a single A40 (48GB) or single L40S GPU. Your GPU should have enough memory to hold DeepFloyd/IF-I-XL-v1.0 and DeepFloyd/IF-II-L-v1.0 models. You should have a CUDA toolkit installed.

Environment

Create or start with a conda environment with python>=3.10,<3.13. To start fresh,

conda env create -n radmesh "python>=3.10,<3.13"

Inspect the script env_setup.sh to change the CUDA version if you need to. Make sure you're in your conda environment, then execute the env_setup.sh script (or run the commands manually yourself) to install dependencies.

Minimal usage

If you just need radmesh/deformations.py to use the geometry processing code, then only numpy, scipy, torch, libigl>=2.6.2, cholespy, and thlog are needed. (See below for a command to install just thlog.)

Note that radmesh/deformations.py depends on radmesh/pytorch3d/, which only requires torch.

(We've vendored just the useful parts of pytorch3d for this, to prevent a bulky installation of the whole pytorch3d package.)

DeepFloyd IF

To set up HuggingFace Hub and your account for downloading the DeepFloyd stages (instructions from DeepFloyd IF):

1. If you do not already have one, create a Hugging Face account
2. Accept the license on the model card of DeepFloyd/IF-I-XL-v1.0
3. Log in to Hugging face locally. In the conda environment you just created, install huggingface_hub

pip install huggingface_hub --upgrade

run the login function in a python shell

from huggingface_hub import login

login()

and enter your Hugging Face Hub access token.

Usage

Preparing your run

Take a look at example-config-localized.json. We've included an example input mesh and prompt in example-run, which you can run immediately with this config file (see Running).

For your inputs: edit paths to the input mesh, vertex selection file, output files, and the prompt.

• The example-config-localized.json file is for runs with a selection region. Deformation and remeshing will only occur within that region. example-config-wholemesh.json is for runs without a selection region, i.e. the whole mesh is allowed to remesh and deform.

The input mesh file can be any format supported by igl.read_triangle_mesh. The vertex selection file, if not null, should be an .npy file of a boolean array of shape (n_verts,), where True indicates the vertex is selected and enabled for deformation and remeshing.

Hyperparameters

The default hyperparameters in the two files should suffice in most cases (example-config-localized.json for localized runs and example-config-wholemesh.json for whole-mesh/global runs). However, inspect the hyperparameters and see if there are any you would like to change before running.

Some adjustments you may wish to consider depending on your mesh:

• If you can, it is recommended to isotropically remesh your input mesh first to have a face count in the 8k-20k range (just so that the average edge length calculation for our default targetlen_schedule settings is a reasonable value, since that schedule is, by default, a multiplier on the initial average edge length). However:

  • Higher-resolution meshes should still work fine, but for best results (since the best overall deformations are achieved with a coarse-to-fine schedule, emphasis on the initial coarseness), some changes to the targetlen_schedule may be applied as needed.
  • Meshes/regions with non-isotropic triangulations (such as a triangulation of a quad-dominant mesh) should also work, since we will do isotropic remeshing to begin with.
  • Meshes with multiple connected components are supported. However, each component must be manifold.

• For human shapes and other tall, slender shapes that take up little volume when normalized to fit the standard cube bounding box, consider using a dist_minmax of [1.4, 2.6] and elev_minmax of [0.0, 30.0].

• If you find that initial inflation in the normal direction of the selection region is too large, try dilating the selection to have a larger area/volume (the volume after hole-closing, which you can see in a printout.) Likewise, if you'd like a larger initial inflation as seed geometry to start optimization with, also consider eroding the selection so that the initial inflation heuristic computes a bigger length.

  • To judge this, check the initialization mesh (after initial inflation + initial remesh). It is saved as a file with the filename pattern drmsh-*-rmsh0-initialization--optm0.npz, which you can view with view_drmsh_npz.py.

  • You can also see it as the first frame of the saved psrec-*.npz recording of a complete optimization run.

• Optimization in general is a little sensitive to the initialization, so if you aren't getting satisfactory results, try changing the above settings related to this initial inflation and your selection region.

Running

Run with:

python run_optimization.py -c example-config-localized.json

(you can override fields on the command line; see python run_optimization.py -h)

Environment variables to make visible to the python process:

• NO_POLYSCOPE=1 (required for headless systems)
• CUDA_HOME pointing to a CUDA toolkit installation matching the CUDA version you installed packages for. nvdiffmodeling will compile an extension module at runtime.

Viewing results

A recording .npz file is saved (e.g. example-run/psrec-spot-wings.npz); play the recording with

thlog replay example-run/psrec-spot-wings.npz

on a system with a monitor (you only need to install thlog and pillow for playback.)

• The command to install thlog with the pillow dependency is

  pip install 'thlog[pil] @ https://github.com/namanhd/thlog/archive/main.zip'

A result .npz file is saved regularly (see save_at_epochs in the config); these files can be viewed with

python view_drmsh_npz.py <drmsh-filename-here>

• (You only need numpy and polyscope, which are also installed when you install thlog.)

Caveats

• We don't fix the seed in general because this pipeline is nondeterministic even with seeding, due to nondeterministic GPU algorithms (e.g. sparse bmm) compounding over the course of thousands of epochs. Set the environment variable TORCH_PLS_BE_DETERMINISTIC=1 for deterministic algorithms, but this comes with a heavy performance penalty.

• Because of this, you can try doing a run several times to see (possibly better) variant results.

BibTeX

@inproceedings{dinh2026radmesh,
  title     = {RADmesh: Remesh-Aware Mesh Deformation},
  author    = {Dinh, Nam Anh and Lang, Itai and Stein, Oded and Hanocka, Rana},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}