DensiCrafter: Physically-Constrained Generation and Fabrication of Self-Supporting Hollow Structures

Shanghai Innovation Institute, Tongji University

The rapid progress of 3D generative models has enabled automatic synthesis of geometry and texture from multimodal inputs (e.g., text or images). However, most existing methods overlook physical constraints and manufacturability, resulting in designs that are difficult to fabricate or structurally unsound.

DensiCrafter addresses this gap by generating lightweight, self-supporting 3D hollow structures through density-field optimization. Starting from coarse voxel grids produced by TRELLIS, we reinterpret the grids as a continuous density field and introduce three differentiable, physically constrained, simulation-free loss terms. A mass regularization term penalizes unnecessary material while a restricted optimization domain preserves the object’s exterior geometry.

Our framework integrates seamlessly with pretrained TRELLIS-based 3D generative models (e.g., TRELLIS, DSO) without architectural modifications. Experiments demonstrate up to 43% reduction in material mass for text-to-3D generation while maintaining high geometric fidelity and improving structural stability. Real-world 3D-printing tests further validate that the produced hollow structures are manufacturable and self-supporting.

---

Dataset

The validation / test dataset can be downloaded from:

🔗 Baidu Drive: https://pan.baidu.com/s/1Wr06oLWo7z7-Ba9h91Zcew?pwd=jtbw

The validation set index file is located at:

./data/val_index

---

Installation & Setup

Install TRELLIS

Please follow the official installation instructions from the 🔗 TRELLIS Repository.

Run

To reproduce our examples for multi-view, single-view, and text-to-3D generation, simply run:

python main.py

Below is the core workflow (you may modify it to fit your own data):

root_path = "./pretrained-models/"

# ---------------------------
# Image-to-3D Pipeline
# ---------------------------
pipeline_image = TrellisImageTo3DPipeline.from_pretrained(
    root_path + "TRELLIS-image-large"
)
pipeline_image.cuda()

image_paths = [
    "./test_200_new/1/1_1.png",
    "./test_200_new/1/1_2.png",
    "./test_200_new/1/1_3.png",
]

run_multi_image_task(pipeline_image, image_paths, "results/multi-image/1")
run_single_image_task(pipeline_image, image_paths[2], "results/single-image/1")

# ---------------------------
# Text-to-3D Pipeline
# ---------------------------
pipeline_text = TrellisTextTo3DPipeline.from_pretrained(
    root_path + "TRELLIS-text-xlarge"
)
pipeline_text.cuda()

prompt = (
    "A stylized cartoon cat character with a spherical head, pointed ears, prominent eyes, a curved mouth, a cylindrical torso with a white patch, elongated arms with large hands, tapered legs, and rounded feet. Predominantly black with white details."
)

run_text_task(pipeline_text, prompt, "results/text/1")

---

Evaluation

1. Dependencies

   • MuJoCo for physics simulation
   • Blender and bpy for mesh rendering

2. Evaluation Code

   • The evaluation scripts are provided in evaluation.py.
   • You may need to modify the get_test_sample function to match your specific task or dataset.

3. Example: Text-Conditioned Evaluation

Below is an example of evaluating a text-conditioned mesh:

# Load a test sample for a text prompt
test_sample = get_test_sample(
    "./results/text/1", 
    cond="A big apple.",  # Text prompt as condition
    trellis_out_path="./results/text/1/final.glb",  # Replace with true trellis result to compute CD/F-score
    id_=1
)

# Evaluate mesh using eval_mesh
res = eval_mesh(
    test_sample["surface"],
    input_condition=test_sample["cond"],
    target=test_sample["target"] if "target" in test_sample else None,
    inner=test_sample["inner"] if "inner" in test_sample else None,
    render=0,  
    device=device,
    cache_dir=sample_cache_dir,
    base_id=test_sample["id"],
    render_mesh=test_sample["texture_mesh"]
)

print(res)
# Excepted return format:
# {'volume': 0.022053713604201896, 'stand': True, 'rotation angle': 12.09500502380615, 'q': array([ 9.94434892e-01, -1.81030881e-03,  3.74934561e-04, -1.05336739e-01]), 'chamfer distance': 0.004543903043337477, 'f score': 0.999549797408834, 'clipscore': 13.835853576660156}

---

Feel free to ask me if you have any questions.

Citation

If you find this repository helpful for your research, please consider citing:

@inproceedings{dang2026densicrafter,
  title={DensiCrafter: Physically-Constrained Generation and Fabrication of Self-Supporting Hollow Structures},
  author={Dang, Shengqi and Chai, Fu and Li, Jiaxin and Yuan, Chao and Ye, Wei and Cao, Nan},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={40},
  number={1},
  pages={193--201},
  year={2026}
}