[zednaked]

Problem

With dense brushes, like most imported Procreate smudges (spacing around 0.04), the smudge tool blurs in place instead of dragging paint. The result is a soft, even-coloured patch with no trace of the brush's shape or grain. Measured offscreen over thin colour stripes with a commercial "Gouache Flow" smudge: the stroke left one flat average colour, and broad colour bands showed only a few pixels of drag.

Three causes, each measured:

1. Pull per stamp. What the brush carries keeps DEFAULT_PULL (0.93) at every stamp. A dense brush stamps many times per pixel, so it lets go of the paint within a fraction of its own size.
2. Deposit saturates. Each stamp lays mix(w, carried, shape·alpha·grain). Hundreds of overlapping stamps take every pixel to 100% whatever the grain, so the texture evens out to a flat colour.
3. RGB mixing. Colours mix as light. Blue dragged into yellow goes grey rather than green.

Changes

• Pull by distance (canvas.rs): pull^(distance / (0.1 · diameter)). Any spacing drags as far as a brush spaced at 10% of its size, the common case, which behaves as before.
• Strongest stamp, not a sum (smudge.wgsl, the amount buffer now always made): a pixel takes the newest carried colour at the strength of the strongest stamp that passed over it. That strength is a crisp cut of shape × grain, so each fibre drags all the way or not at all. The pixel as the stroke found it is kept in amount's second half.
• Opacity curve: the strength sets how much is laid and how far it drags, both at alpha^0.15. Measured over the slider, the effect now rises evenly from 10% to 100%: 0.40, 0.53, 0.68, 0.84, 1.03, 1.25, 1.52, 1.86, 2.30, 2.91. Before, almost nothing happened up to 50%, the effect jumped 10× between 50 and 70%, and it was flat above.
• Pigment mixing in the carry and the deposit: Kubelka–Munk over 38 spectral bands. The tables are generated from spectral.js 3.0.0 (MIT, © 2025 Ronald van Wijnen), credited in the shader. Shares are linear rather than squared as in spectral.js, because squared shares also lengthened the drag. The inverse 1 + ks − √(ks² + 2ks) is written as 1 / (1 + ks + √(ks² + 2ks)), so it doesn't lose precision in 32 bits.

Checked

• New test a_smudge_mixes_blue_into_yellow_as_paint_does: it fails on main (the greenest pixel is −3, an even grey) and passes here.
• cargo test -p procraft-gpu: 174 passed, 0 failed, including a_smudge_goes_by_a_shape_at_its_own_size and a_stroke_reads_a_brush_file_s_grain_at_its_own_size. Both need the crisp cut's floor of 0.12, without which a zero in the shape or grain passed at half strength.
• Used for painting on desktop Linux (Wayland, Kamvas pen) by an artist who uses Procreate daily.

Open questions

• At full strength the smudge picks up what it has just laid, so a strong drag runs to the end of the stroke. Reading the pick-up from the layer as it was before the stroke (source_before) didn't make it finite in my test, and I couldn't explain why, so I left it out.
• I couldn't compare against Procreate directly. The calibration (0.1 of the diameter, the 0.12 cut, the 0.15 exponent) was tuned by feel with the artist.

🤖 Generated with Claude Code