DVD: Discrete Voxel Diffusion for 3D Generation and Editing

This repository hosts the pretrained DVD checkpoints for Discrete Voxel Diffusion for 3D Generation and Editing.

DVD is a discrete diffusion pipeline that generates or edits a pure 64^3 voxel grid. The voxel output can be used directly, exported as a cubified voxel mesh, or passed as the sparse-structure anchor for TRELLIS stage 2 to produce final mesh and 3D Gaussian assets.

• Paper: arXiv:2605.07971
• Spaces: DVD Image, DVD Text

Checkpoints

Variant	Files	Use case
Image generation	dvd_img.json, dvd_img.safetensors	image-conditioned voxel generation
Image editing	dvd_img_BSP_ft.json, dvd_img_BSP_ft.safetensors	image-conditioned voxel editing
Text generation	dvd_text.json, dvd_text.safetensors	text-conditioned voxel generation
Text editing	dvd_text_BSP_ft.json, dvd_text_BSP_ft.safetensors	text-conditioned voxel editing

Minimal Usage

from dvd import DVDImageToVoxelPipeline, DVDTextToVoxelPipeline

repo_id = "Zhengrui/dvd"

image_dvd = DVDImageToVoxelPipeline.from_pretrained(repo_id, variant="base", device="cuda")
image_edit_dvd = DVDImageToVoxelPipeline.from_pretrained(repo_id, variant="bsp", device="cuda")

text_dvd = DVDTextToVoxelPipeline.from_pretrained(repo_id, variant="base", device="cuda")
text_edit_dvd = DVDTextToVoxelPipeline.from_pretrained(repo_id, variant="bsp", device="cuda")

Example image-conditioned generation:

from PIL import Image
from dvd import DVDImageToVoxelPipeline, export_cubified_voxels

image = Image.open("assets/example_image/T.png")
dvd = DVDImageToVoxelPipeline.from_pretrained("Zhengrui/dvd", variant="base", device="cuda")
voxels = dvd.sample_voxels(image, seed=42, steps=256)
export_cubified_voxels(voxels, "dvd_voxels.glb")

Voxel Convention

DVD stores voxel outputs as:

samples: [B, D, H, W]
coords:  [N, 4] as [batch, x, y, z]

Saved .npy coordinate files usually omit the batch dimension and can be loaded back with as_voxel_output(coords, resolution=64).

Notes

These files are the DVD voxel-stage checkpoints. TRELLIS stage-2 checkpoints are loaded separately from the official TRELLIS model repositories, such as microsoft/TRELLIS-image-large and microsoft/TRELLIS-text-large.

Citation

@misc{xiang2026dvddiscretevoxeldiffusion,
      title={DVD: Discrete Voxel Diffusion for 3D Generation and Editing},
      author={Zhengrui Xiang and Jiaqi Wu and Fupeng Sun and Heliang Zheng and Yingzhen Li},
      year={2026},
      eprint={2605.07971},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.07971},
}