Fractal Gen Cuda

FractalGeneratorV3 ("Fractal Gen Cuda" in the UI) renders actual fractals - Mandelbrot, Julia, Burning Ship, Tricorn, Newton - as images inside your workflow. No model, no diffusion: just iteration math, and it's fast because the compute runs as tensor ops on the GPU via torch, falling back to CPU when there's no CUDA. It's the texture/background generator you didn't know you needed: want a trippy backdrop, a lathe-ready texture, or a source image to img2img a "cosmic" render from? Generate the fractal, feed it to the sampler, done.

What it generates

The preset dropdown picks the fractal family:

• Classic Mandelbrot - the default, and the one people mean when they say fractal.
• Julia Set - same math, but the iteration constant is free, so it produces those swirling, seashell-ish shapes.
• Burning Ship - Mandelbrot with absolute values in the formula; looks like a burning vessel when zoomed, hence the name.
• Tricorn - the Mandelbar, its symmetry gives it a repeating three-lobed look.
• Newton - a Newton fractal based on complex roots; sharp, geometric, and the one that benefits most from GPU speed because it's the most iteration-hungry.

zoom_level (0.1–100) magnifies into the set; at high zooms you're deep in detail, so raise max_iterations (50–2000, default 500) or the deep areas go blurry. power (2–5) changes the exponent in the iteration formula - even small changes transform the pattern completely.

The inputs that matter

• width / height (64–4096, default 1024) - output resolution. This is where CUDA pays for itself; 2048+ renders are where a CPU fallback gets painful.
• color_preset - eight schemes from "Classic White-Grey" through "Fire", "Deep Space", "Psychedelic". Combined with color_cycles (0.1–10) which repeats the gradient for more color bands.
• smooth_coloring - on by default, and worth keeping on; it removes the hard iteration banding so colors flow instead of step.
• x_center / y_center (optional) - pan the view. Only really needed with the Custom preset, where you pick the coordinates and the iteration details yourself. The named presets set their own centers.

Output is a single IMAGE. Nothing else - you wire it into your sampler, VAE, or image nodes like any source image.

How it works

Under the hood it vectorizes the escape-time algorithm over a torch grid, iterating with abs(Z) <= escape_radius masks, then smooth-colors by log-scaled iteration counts. All the color mapping and fractal-family math happens in the node - you can't get "the Mandelbrot set as a latent," because it renders pixels first. That's the right call; the IMAGE output plugs into any ComfyUI pipeline.

The honest take

Is it a must-have? No - you could generate fractals in any fractal program and import them. But as a procedural source inside a workflow it's genuinely nice: no file management, fully parameterized, deterministic per seed-of-input, and you can chain dozens of renders by feeding the image into a ControlNet or img2img pass to turn it into something that looks nothing like a fractal at all. Start with Classic Mandelbrot at 1024² and 500 iterations, get the hang of zoom, then let the Newton preset melt your GPU.