🌊 ComfyUIFlowFields

This node makes the kind of image you'd normally fight a particle sim all afternoon for: swirling, ink-in-water trail lines, in under a second. One node, no model, no VRAM. It builds a flow field - a mathematical "wind map" across the canvas - drops particles in, and renders where they get blown. And the name is a lie in the best way: no API, no key, and its only link to AI is that it produces conditioning you can hand to the real model.

It's from ComfyUIFlowFields, a solo pack by javierjarart (Javier Jara), MIT-licensed and in the Comfy Registry. Small, fresh, README in Spanish, but well built - clean core//nodes/ split, a real pytest suite, and one bugfix you wouldn't expect from a weekend node (more below). It ships a single node, FlowFieldGenerator, under the Jarart Tools category.

How it works

The core is curl noise, a divergence-free vector field - think incompressible fluid, so nothing pools or vanishes; particles follow a coherent swirl instead of drifting into a dead spot. The field is a fractal sum of sinusoids (noise_octaves sets the fBm layers), and the flow direction comes from the gradient of that scalar noise.

Particles spawn, then advect step by step: each samples the field via bilinear interpolation, moves along it, and wraps at the canvas edges like a torus. Every step draws a short segment per particle onto a fading canvas - that's the trails - while a separate accumulator tracks raw density for the mask. The whole sim is numpy + Pillow on CPU; torch only appears at the end to convert arrays to tensors.

The inputs that actually matter

Fourteen inputs, and you'll touch maybe five. The rest are fine at defaults:

• noise_scale - the big one. Lower = broader, smoother swirls; higher = tighter, busier turbulence. Default 0.008 is a good starting point for organic looks.
• steps - more steps = longer trails. Default 150 is plenty for a poster; the trails get denser and more tangled as you climb.
• fade - how fast trails persist before the background wipes them. Lower fades give thin current-like streaks; higher builds up thick glowing rivers.
• spawn_mode - where particles start: random, center, grid, or circle. center with a low fade gives a nice radial burst; grid is a uniform texture.
• flow_angle - rotates the entire field 0–360°, handy for a diagonal sweep.
• seed - reproducibility, natch.

color_mode (rainbow, fire, ocean, monochrome, custom) and bg_color_hex are cosmetic; width/height, particle_count, speed, and line_width do what they say.

The three outputs

• particle_image - the RGB trails. Save it or preview it, or use it as an IP-Adapter style reference on an SD1.5/SDXL workflow (that adapter path mostly isn't there on modern Flux-era bases).
• density_mask - a soft MASK of everywhere particles passed. The sneaky-useful one: wire it into Set Latent Noise Mask before a KSampler and sampling focuses on the swirl regions, the mask-driven trick from the KB's detailing playbook. Preview it via MaskToImage.
• flow_map - a color-coded map of the field (R = X direction, G = Y direction, B = magnitude). It's a guide image, not a true ControlNet preprocessor output - treat it as a visual reference, not plug-and-go.

Installing it

Two ways, both painless:

# ComfyUI Manager: search "ComfyUIFlowFields" and install
# or, by hand:
cd ComfyUI/custom_nodes && git clone https://github.com/javierjarart/ComfyUIFlowFields

Restart and done. Dependencies are numpy, Pillow, and torch - all already in ComfyUI, so no model downloads, no heavy installs. That's the whole appeal: the anti-node of a 5GB checkpoint.

Gotchas worth knowing

• bg_color_hex expects exactly six hex digits, no #. Malformed values (like the #0A0A14 you'll instinctively type) get rejected before the sim runs with a clear error, not a mid-render crash.
• The seed trap is already fixed. ComfyUI's randomize button rolls 64-bit seeds; numpy's RandomState only accepts 32-bit, a recipe for crashes. This pack masks seeds down to 32 bits, so any seed works - just know a seed and its value plus 2³² produce the same image.
• It's a CPU sim, so respect the sliders. Each step draws one line per particle in a Python loop. Defaults (800 × 150) run in a couple of seconds; 5000 particles × 1000 steps is tea time. Keep steps modest unless you're chasing long trails.
• Particles wrap around the canvas edges (torus world). The renderer skips the wrap jump so you don't get a stray long streak - but particles do teleport to the far side, which shapes the composition in surprising ways.

If you want organic background texture, or a mask to steer a masked sampling pass, this is the one to reach for. It won't change your life, but it costs nothing to run and it's genuinely fun to dial in.