Field Scatter

Random-looking dots that are actually a grid

Field Scatter places one shape per lattice cell - a circle, rectangle, polygon or star - with seeded jitter on position, size, rotation and value. It's the node for confetti, bokeh, dust motes, painted speckle, a field of soft lights, or a stipple mask you want to composite an effect through. And it's the deterministic kind of random: same settings, same seed, same stamps, every run, at every resolution.

That last part is the pack's whole thesis applied to scattering. density is cells across the reference dimension S = max(H, W), so cells are square at every aspect ratio - the same scatter at 512 and at 2048, just resolved more finely.

The controls worth understanding

• shape, sides, star_ratio, stamp_aspect - the stamp geometry. stamp_aspect is rect-only and controls the native half-width, so an oblong stamp is exact rather than a stretched circle.
• density - cells across S. Default 8. This is your "how many slots are there" control.
• size - the stamp radius in cell units, where 1.0 is one full cell width. That's an unusually nice choice: size means the same thing whether you've got 8 cells or 40, so changing density doesn't also change how big the dots are.
• fill - the probability any given cell is occupied. 1.0 means every cell stamped, which gives you a perfect grid rather than a scatter. Drop it to 0.4 and you get a scatter. This is the knob that makes it look random.
• position_jitter, size_jitter, rotation_jitter, value_jitter, rotation - each jitter draws from an independent per-cell hash channel driven by seed. That independence matters: turning up size jitter never reshuffles which cells are occupied or how they're rotated. You can tune one thing without losing the arrangement you liked.
• falloff and aa_width - the authored soft edge and the antialiasing width, exactly as on Field Shape. With falloff at 0 you get hard stamps.
• distribution, coverage, invert, seed, width/height - the standard pack tail.

size is clamped by the 3×3-neighbourhood reach cap (a stamp can only overhang so far before it would need to read cells it doesn't have), and star_ratio is clamped to what's achievable for the chosen sides. Both cases print a console note. If your stamps look cut down from what you asked for, that's the cap, not a bug.

The optional inputs are the good stuff

• reference_image / reference_mask - inherit H, W, batch and device. reference_mask is size-only; it supplies dimensions and never touches the scatter itself. People wire a mask here expecting it to control placement and then wonder why the dots are everywhere. It won't. Use fill_mask for that.
• fill_mask - spatially modulates placement probability. Read once per lattice cell at the cell's centre, so effective fill = fill × (mask_min + (1 − mask_min) · m^mask_gamma). White means full fill; black means mask_min × fill.
• mask_min - the placement floor where fill_mask is black. 0 thins to nothing; raise it to keep some stamps everywhere. Only active when fill_mask is wired.
• mask_gamma - a contrast curve on the mask before the floor. 1.0 skips the operation entirely.

Two behaviours to know about fill_mask. Stamps land whole or not at all - edges are honoured at cell resolution, and a stamp can overhang by its own reach - so keep mask features at least two cells wide or you'll get a ragged boundary. And with distribution=uniform, the area that gets masked out is remapped to mid-grey rather than black, because the rank transform ranks the whole frame. If you need truly zero outside the region, use native.

Outputs are mask and preview.

Install

Manager → ComfyUI Field, or:

cd ComfyUI/custom_nodes
git clone https://github.com/jeremieLouvaert/ComfyUI-Field

Restart. No models, no extra packages. Under AKURATE/Fields/Generate.

Getting burned

fill at its default of 1.0 stamps every cell, which reads as a grid, not a scatter. The first thing to do with this node is drop fill to about 0.5 and add a little position jitter - that's when it starts to look organic.

Also: when the output is provably binary (falloff 0, value_jitter 0, aa_width 0), the node forces native regardless of what distribution says. The rank transform of a two-valued image isn't useful, so the pack just refuses. If the distribution widget seems to do nothing on hard stamps, that's why.

Finally, the seed widget is only active when it matters: 0 < fill < 1, any jitter above zero, or fill_mask wired. With fill = 1 and no jitter, changing the seed genuinely does nothing, and that's by design rather than a stuck widget.