DVD for ComfyUI

DVD: Deterministic Video Depth Estimation with Generative Priors (EnVision Research, HKUST(GZ)) in ComfyUI's file layout, for ComfyUI-DVDepth.

DVD is Wan 2.1 T2V 1.3B with a rank-512 LoRA, run as a regressor: clean video latents in at a fixed timestep, depth latents out, one forward pass. The files here are the released checkpoints (FayeHongfeiZhang/DVD) with the LoRA merged into the base weights (in fp32, then stored as bf16) and the diffsynth prefixes removed, so ComfyUI's Load Diffusion Model detects them as a plain Wan 2.1 T2V model. Nothing else changed; the ComfyUI path matches the reference implementation to 1e-6 in fp32.

File	Folder	Source checkpoint
diffusion_models/dvd_1.1_bf16.safetensors	models/diffusion_models	dvd_1.1.safetensors (DVD v1.1, April 2026: more robust to reflective surfaces, better temporal consistency)
diffusion_models/dvd_1.0_bf16.safetensors	models/diffusion_models	model.safetensors (DVD v1.0)

The VAE is the stock Wan 2.1 VAE: use Comfy-Org/Wan_2.1_ComfyUI_repackaged wan_2.1_vae.safetensors (the copy inside the DVD checkpoints is bit-identical). No text encoder is needed: DVD conditions on an all-zero text context.

Made with scripts/convert.py from the node pack.

License

The DVD weights are released under CC BY-NC 4.0 by their authors (non-commercial use only); these files are a format conversion of them and carry the same license. Wan 2.1 is Apache-2.0.

@article{zhang2026dvd,
  title={DVD: Deterministic Video Depth Estimation with Generative Priors},
  author={Zhang, Hongfei and Chen, Harold Haodong and Liao, Chenfei and He, Jing and Zhang, Zixin and Li, Haodong and Liang, Yihao and Chen, Kanghao and Ren, Bin and Zheng, Xu and others},
  journal={arXiv preprint arXiv:2603.12250},
  year={2026}
}