NormalCrafter (`normal_crafter`)

Sequence · Surface normal map estimation (diffusion video prior)

Estimate per-pixel surface normals across a clip, encoded as RGB where the channels map to the X/Y/Z components of the unit normal vector. The plugin loads a contiguous window of frames into ComfyUI in one job so the upstream SVD-based sliding-window sampler can produce temporally coherent normals across time.

What you give it

An RGB clip.

What you get back

A per-frame surface normal map (RGB-encoded unit normals).

Typical VFX uses: relighting and synthetic shading passes, normal-based AOVs for matte/roto refinement, retopology and projection-mapping guides, bump and displacement extraction, re-shading of plate elements without rebuilding geometry.

Commercial use

Component	License	Commercial OK?
Code (`Binyr/NormalCrafter`)	MIT	✅ Yes
NormalCrafter weights	Apache 2.0	✅ Yes (in isolation)
Stable Video Diffusion base (pulled at runtime)	Stability AI Non-Commercial Community License	❌ No

The runtime SVD dependency is the binding constraint. Not safe for paid production without engaging Stability AI for a commercial license.

Requirements

GPU VRAM: ~20 GB at 1024×576, ~6 GB at 512×256.
ComfyUI custom node: AIWarper/ComfyUI-NormalCrafterWrapper.
Model weights: the shipped workflow has no separate model-loader node — the NormalCrafterNode custom node resolves and downloads its (SVD-based) weights internally on first use, including the Stable Video Diffusion base it builds on. Upstream source: Yanrui95/NormalCrafter (SVD-XT base: stabilityai/stable-video-diffusion-img2vid-xt, ~9 GB, pulled on first run).
Resolution constraint: dimensions are typically rounded to multiples of 64 (SVD constraint).

Parameters

Parameter	Meaning
Seed	Random seed for diffusion sampling. Change to vary the result; keep the same value for reproducible runs.
Max Res Dimension	Maximum resolution along the longest image side (`max_res_dimension`). Larger values give finer detail but require more VRAM; 1024 is the recommended default.
Window Size	Number of frames processed per temporal window. Must not exceed the total frame count (clamped down to Frame Limit automatically). Larger windows improve temporal consistency but require more VRAM.
Time Step Size	Diffusion time-step size. Higher values are faster but lower quality.
Decode Chunk Size	Number of frames the VAE decodes at once. Lower values use less VRAM at the cost of speed.
Frame Limit	Maximum number of frames loaded from the sequence (`image_load_cap`). Set to 0 for no limit. Keep ≤ Window Size for best results.
Offload Pipe On Finish	Move the NormalCrafter pipeline to CPU once inference completes, freeing VRAM for downstream nodes.
Use xFormers	Memory-efficient attention via xFormers. Choices: Auto (use it if available), Enable (force on), Disable (force off).
Wrapper-disabled knobs	`fps_for_time_ids`, `motion_bucket_id`, `noise_aug_strength` were observed by the wrapper author to have minimal effect and are hardcoded — not exposed as parameters.

Plus the standard ComfyUI base parameters (server URL, mount paths, project name, workflow path).

Demos & comparisons

NormalCrafter pipeline: video-to-normal-map diffusion architecture. NormalCrafter pipeline diagram. © Bin et al., ICCV 2025. Source: normalcrafter.github.io. Used for documentation under MIT attribution.

Input → output

Input clip and the NormalCrafter normal-map result (upstream side-by-side reel)

Comparison reel. © Bin et al., ICCV 2025, normalcrafter.github.io. Reproduced under MIT attribution; cite arXiv:2504.11427.

To see what this model produces, the upstream sources have the most authoritative demos:

Project page — normalcrafter.github.io — side-by-side video comparisons against per-frame normal estimators.
GitHub README — Binyr/NormalCrafter — open-world clips with their normal sequences.

Image attribution: Bin et al., ICCV 2025; reproduced for documentation purposes with citation to arXiv:2504.11427. See the credits page.

Limitations

Window-scale drift: on very long clips processed in multiple sliding windows, scale/orientation drift can occur at window joins. Use a larger Image Load Cap when VRAM allows so the whole shot fits in one pass.
Failure modes: transparent / refractive surfaces, mirrors, strong speculars, very low-light footage with crushed blacks, fine subpixel structures (hair, foliage edges), and heavy motion blur all degrade output.
Output convention: RGB-packed unit normals; verify camera vs world space and axis convention in your shading pipeline before relying on the values.

Credits

Yanrui Bin, Wenbo Hu, Haoyuan Wang, Xinya Chen, Bing Wang. NormalCrafter: Learning Temporally Consistent Normals from Video Diffusion Priors. ICCV

Paper · Project page · GitHub

ComfyUI wrapper by AIWarper.

Citation

@inproceedings{bin2025normalcrafter,
  author    = {Bin, Yanrui and Hu, Wenbo and Wang, Haoyuan and Chen, Xinya and Wang, Bing},
  title     = {NormalCrafter: Learning Temporally Consistent Normals from Video Diffusion Priors},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2025},
  eprint    = {2504.11427},
  archivePrefix = {arXiv}
}

NormalCrafter (normal_crafter)