(deforum) Add Advanced Noise

(deforum) Add Advanced Noise is the sibling of (deforum) Add Custom Noise for people who got bored of gaussian. Where the basic node offers 19 practical flavors, this one serves the exotic menu: wavelet, value, flow, turbulence, ridged_multifractal, reaction_diffusion, voronoi, and simplex. These are the noise types you'd find in a graphics-toolkit textbook - the kind that make water surfaces, rocky terrain, and cellular patterns. If your animation needs a texture that behaves like a physical pattern rather than just static, this is the node.

How it works

Same shape as its sibling: input images, pick a noise_type, set amount, and it returns both the noised image and the raw noise map. The difference is the math behind each type. Some highlights:

• wavelet - decomposes noise via wavelet transforms (haar, db1, sym2, coif1, bior1.3, rbio1.3 - pick from the wavelet dropdown) with a chosen mode (symmetric/periodic/reflect/zero-padding). Gives you band-limited, controllable-frequency noise.
• voronoi - cellular noise from num_points scattered seed points. Great for organic cells, cracked patterns, scales.
• simplex - smooth gradient noise (via the opensimplex library); scale, octaves, persistence and lacunarity shape its fractal structure.
• reaction_diffusion - literally simulates the Gray-Scott reaction-diffusion model (Du, Dv, feed_rate, kill_rate, steps). This is the one that produces those mesmerizing spot-and-stripe Turing patterns. It's slow, because it's an actual simulation, but it's also the one nothing else in the pack can do.
• flow / turbulence / ridged_multifractal - directional and fractal noise families; frequency, theta, sigma_x, sigma_y, flow_scale, flow_angle steer them.

The inputs and outputs that matter

• images - the batch to corrupt.
• noise_type - the 8-type menu above.
• amount (FLOAT, 0–100, default 0.1) - global intensity.
• seed (optional) - lock it for reproducible patterns.

Everything else is per-type tuning: num_points (voronoi), scale/octaves/persistence/lacunarity (fractal family), wavelet/mode (wavelet), frequency/theta/sigma_x/sigma_y/flow_scale/flow_angle (flow/gabor-ish), and the reaction-diffusion quartet Du, Dv, feed_rate, kill_rate, steps, plus res. You'll mostly leave the irrelevant ones alone - the node ignores parameters the chosen type doesn't use.

Outputs:

• NOISED_IMAGE - the image with the pattern applied.
• NOISE - the raw noise/pattern map, so you can composite it yourself.

Installing this pack

Part of Deforum Nodes by XmYx. ComfyUI Manager: search Deforum Nodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/XmYx/deforum-comfy-nodes

Restart, let install.py pull the deforum-studio backend from GitHub plus the scientific-Python stack (numpy < 2.0, Python 3.10 required). The pywt (PyWavelets) and opensimplex libraries come along for the ride - if you get an import error on the wavelet or simplex types, those two are the missing pieces and are normally installed by the pack's installer.

Common issues

• Reaction-diffusion crawls. It's a real simulation at full image resolution. Use a small res/steps, generate the pattern, then upscale - don't run it at 1080p directly.
• Import errors on wavelet or simplex. Missing pywt / opensimplex. If the pack's installer didn't complete, pip install pywt opensimplex in your ComfyUI venv fixes it.
• Noise is there but the pattern is invisible. amount at 0.1 on a low-contrast pattern is easy to miss. Raise it; these advanced types generally need more headroom than plain gaussian.