DepthCrafter (depth_crafter)

Sequence · Temporally consistent video depth (diffusion)

A video diffusion model that estimates a depth sequence with temporal consistency over an entire window of frames jointly, rather than estimating each frame independently.

What you give it

• An RGB clip.

What you get back

• A depth sequence, one depth map per input frame, scaled in the same relative-depth space across the whole window. No flicker, no per-frame swim.

Typical compositing uses: 2.5D camera moves, depth-based defocus and atmospherics, relighting, rotoscoping assists, stereo / parallax generation from monocular plates — anywhere per-frame methods would create a wobble that sells the depth pass as fake.

Commercial use

Component	License	Commercial OK?
DepthCrafter code (Tencent)	Custom Tencent non-commercial	❌ No
DepthCrafter weights	Same Tencent non-commercial terms	❌ No
Stable Video Diffusion XT base	Stability AI Non-Commercial Community License	❌ No

Not safe for paid production work without engaging both Tencent and Stability AI for commercial agreements.

Requirements

• GPU VRAM: ~26 GB at 1024×576, ~9 GB at 512×256. The ComfyUI port supports CPU offload down to an 8 GB minimum at low resolution.
• ComfyUI custom node: akatz-ai/ComfyUI-DepthCrafter-Nodes.
• Model weights: the embedded workflow’s DownloadAndLoadDepthCrafterModel node auto-downloads the weights on first use into the appropriate ComfyUI/models/ subfolder — no manual placement required.
  • tencent/DepthCrafter — ~3.05 GB.
  • SVD-XT base: stabilityai/stable-video-diffusion-img2vid-xt — ~9 GB.
• Resolution: dimensions must be multiples of 64.

Parameters

Parameter	Meaning
Force Size	When enabled, the DepthCrafter node uses its internal fixed resolution; resize source footage upstream to the desired resolution before processing. Default: on.
Inference Steps	Number of diffusion denoising steps. More steps = higher quality but slower. 10 is a good balance; 5 for quick previews. Default: 10.
Guidance Scale	Classifier-free guidance scale. Higher values follow the depth prior more strongly. Default: 1.2.
Window Size	Number of frames processed per temporal window. Must not exceed the total frame count; larger windows improve temporal consistency but need more VRAM. Default: 25.
Window Overlap	Number of overlapping frames between consecutive temporal windows. Higher overlap improves temporal consistency on longer sequences. Default: 25.
Frame Limit	Maximum number of frames to load from the sequence (the model runs all loaded frames jointly). Set to 0 for no limit; keep ≤ Window Size for best results. Default: 25.
CPU Offload	Offload model weights to CPU between inference passes to reduce VRAM. Recommended for GPUs with < 16 GB VRAM. Default: on.
Sequential CPU Offload	More aggressive layer-by-layer CPU offloading; minimises VRAM at the cost of speed. Only enable if CPU Offload alone is insufficient. Default: off.

Plus the standard ComfyUI base parameters.

Demos & comparisons

DepthCrafter overview figure. © Hu et al., Tencent AI Lab, CVPR 2025 Highlight. Source: depthcrafter.github.io. Reproduced under fair-use citation; Tencent code/weights non-commercial.

Input → output

Input frame

DepthCrafter depth output

Side-by-side: input RGB frame and DepthCrafter’s per-pixel depth result. © Hu et al., Tencent AI Lab, CVPR 2025; reproduced under fair-use citation.

The strongest argument for a temporally-consistent depth model is visual. The upstream sources show side-by-side reels against per-frame methods:

• Project page — depthcrafter.github.io — hero comparison videos vs. Depth Anything, Marigold, NVDS, ChronoDepth.
• GitHub README — Tencent/DepthCrafter — animated open-world clips with their depth sequences.

Image attribution: Hu et al., Tencent AI Lab, CVPR 2025 Highlight; reproduced for documentation purposes with citation to arXiv:2409.02095. See the credits page.

Performance

• ~2.1 fps at 1024×576 on an A100 (~465 ms/frame), times the number of diffusion steps.
• ~8.6 fps at 512×256 on the same hardware.
• Quadratic scaling in resolution; sub-linear in window size up to the cap.

Limitations

• Affine-invariant relative depth. Not metric. Don’t interpret raw values as world-scale distances.
• Window scale drift: on very long clips (multiple sliding windows), scale/shift can drift between windows. The plugin and node blend in the overlap region, but extreme cases need post-processing.
• Fast motion: very fast motion or heavy motion blur degrades temporal stability.
• Transparent / reflective surfaces: poorly handled (inherited from the SVD prior).

When to use this vs Depth Anything V3

• Use DepthCrafter when temporal stability dominates the project — e.g. parallax moves, depth-based comp work where flicker is unacceptable.
• Use Depth Anything V3 when per-frame fidelity matters more, when you need metric depth, when you need commercial-license-safe output, or when VRAM and time budgets rule out diffusion.

Credits

Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, Ying Shan. DepthCrafter: Generating Consistent Long Depth Sequences for Open-world Videos. CVPR 2025 (Highlight). Tencent AI Lab. Paper · Project page · GitHub

ComfyUI port by akatz-ai.

Citation

@inproceedings{hu2025depthcrafter,
  author    = {Hu, Wenbo and Gao, Xiangjun and Li, Xiaoyu and Zhao, Sijie and
               Cun, Xiaodong and Zhang, Yong and Quan, Long and Shan, Ying},
  title     = {DepthCrafter: Generating Consistent Long Depth Sequences for Open-world Videos},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2025},
  note      = {Highlight},
  eprint    = {2409.02095},
  archivePrefix = {arXiv}
}