zhaxie/3Deditformer

Towards Scalable and Consistent 3D Editing

👋 Hi, I’m Ruihao Xia, a Ph.D. candidate (expected 2026). I’m seeking internship and full-time opportunities in AIGC, 3D vision, and multimodal intelligence. More about me and my CV: https://xiarho.github.io/ — feel free to reach out if my background aligns with your team!

In this paper, we introduce 3DEditVerse, the largest paired 3D editing benchmark, and propose 3DEditFormer, a mask-free transformer enabling precise, consistent, and scalable 3D edits.

:sun_with_face: 3DEditVerse

Our 3DEditVerse, the largest paired 3D editing benchmark to date, comprising 116,309 high-quality training pairs and 1,500 curated test pairs.

:sparkles: 3DEditFormer

Our 3DEditFormer, a 3D-structure-preserving conditional transformer, enabling precise and consistent edits without requiring auxiliary 3D masks.

:hammer_and_wrench: Environment Setup

1. Our environment setup follows the official TRELLIS project.
   Please refer to their installation instructions for dependency versions and CUDA/PyTorch configurations.

2. Install the blender: Download from https://download.blender.org/release/Blender4.4/blender-4.4.3-linux-x64.tar.xz and extract it.

:nut_and_bolt: Preparing the Datasets

1. Download our 3DEditVerse dataset: 3DEditVerse. About 227 GB (636,569 files).

2. Extract the *.tar files in the 3DEditVerse folder.

tar -xf alpaca.tar / mixamo.tar / test_data.tar

• For flux_edit.part.tar.* files, you should concatenate them into a single file before extracting.

cat flux_edit.part.tar.* > flux_edit.tar

3. The data folder structure should look like this:

path_to_3DEditVerse/3DEditVerse
├── alpaca
│   ├── 1
│   ├── 2
│   ├── ...
├── flux_edit
│   ├── 3D CG rendering_4
│   ├── 3D CG rendering_5
│   ├── ...
├── mixamo
│   ├── latents
│   ├── renders_cond
│   ├── ss_latents
├── test_data
│   ├── alpaca
│   ├── alpaca_render
│   ├── flux_edit
│   ├── flux_edit_render
│   ├── mixamo
│   ├── mixamo_render
├── alpaca_confidence.json
├── flux_edit_confidence.json
├── dataset_info.json
├── test_data_info.json
├── edit_prompts.json

:arrow_forward: Inference and Evaluation with our Trained 3DEditFormer

1. Download the trained model of 3DEditFormer and put them in the ./work_dirs/Editing_Training folder. Then, you can inference on the testing data in 3DEditVerse:

CUDA_VISIBLE_DEVICES=0 python eval_3d_editing.py --cuda_idx 0 --world_size 1 --rank 0 --dataset_root_dir /path_to_3DEditVerse/3DEditVerse --blender_path /path_to_blender/blender-4.4.3-linux-x64/blender --ss_latents_load_id img_to_voxel --latents_load_id voxel_to_texture --save_name 3DEditFormer --output_mesh --output_video --print_time

• In the above command, replace /path_to_3DEditVerse/3DEditVerse with the path to your 3DEditVerse dataset and /path_to_blender/blender-4.4.3-linux-x64/blender with the path to your blender. CUDA_VISIBLE_DEVICES=0 means the GPU index for model inference, --cuda_idx 0 means the GPU index for image rendering with blender.
• You can change the --world_size and --rank to inference the model on multiple GPUs, i.e., run the command with the same --world_size 4 and different --rank 0/1/2/3 on 4 GPUs.

2. Calculate the 2D metrics based on the rendered images (rendered from predicted 3D meshes):

CUDA_VISIBLE_DEVICES=0 python calculate_metric_2d.py --eval_results_dir ./work_dirs/eval_results/3DEditFormer --dataset_root_dir /path_to_3DEditVerse/3DEditVerse

• The metrics will be saved in ./work_dirs/eval_results/3DEditFormer/eval_metric.json.

3. Calculate the 3D metrics based on the predicted 3D meshes:

CUDA_VISIBLE_DEVICES=0 python calculate_metric_3d.py --eval_results_dir ./work_dirs/eval_results/3DEditFormer --dataset_root_dir /path_to_3DEditVerse/3DEditVerse

• The metrics will be saved in ./work_dirs/eval_results/3DEditFormer/eval_metric.json.

:desert_island: Training 3DEditFormer with our 3DEditVerse

1. The first stage: generation of coarse voxelized shapes

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node=4 --master_port=12349 train_torchrun.py --config configs/editing/ss_flow_img_dit_L_16l8_fp16.json --data_dir /path_to_3DEditVerse/3DEditVerse --output_dir ./work_dirs/Editing_Training/img_to_voxel_01 --random_cond_gt --train_only_editing_weights --lr 0.0001 --max_steps 40000 --batch_size_per_gpu 4 --random_ori_edit 0.15 --simple_edit_data_if_filtered

2. The second stage: generation of fine-grained texture

CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node=4 --master_port=12349 train_torchrun.py --config configs/editing/slat_flow_img_dit_L_64l8p2_fp16.json --data_dir /path_to_3DEditVerse/3DEditVerse --output_dir ./work_dirs/Editing_Training/voxel_to_texture_01 --random_cond_gt --train_only_editing_weights --lr 0.0001 --max_steps 40000 --batch_size_per_gpu 4

:label: TODO

• Interactive 3D editing demo.
• Visualize the 3DEditVerse dataset.

:hearts: Acknowledgements

Thanks TRELLIS, VoxHammer for their public code and released models.

:black_nib: Citation

If you find this project useful, please consider citing:

@article{3DEditFormer,
  title={Towards Scalable and Consistent 3D Editing},
  author={Xia, Ruihao and Tang, Yang and Zhou, Pan},
  journal={arXiv:2510.02994},
  year={2025}
}