Track 8 — Shipping

You can now write a preset from nothing. This track covers what happens after: knowing where it will and won't run correctly, keeping it fast, and publishing it.

Compatibility matrix

MilkDrop presets travel between five active engines — the original Winamp MilkDrop 2, projectM, Butterchurn, and Stims on both its WebGL and WebGPU backends — and until now, no single place documented where a given feature works across all of them. This is that matrix, built directly from the Stims compiler's own compatibility/parity code and shader-support inventory.

Read the labels carefully. Cells for Stims are sourced from code (compatibility.ts, parity.ts, shader-execution-classification.ts, the shader support inventory) and are exact. Cells for Winamp MilkDrop 2, projectM, and Butterchurn are marked [inferred] — reasoned from the format spec and public documentation, not verified by running those engines. Treat an inferred cell as "should work, unconfirmed," not as tested fact.

Registers, buffers, control flow

Shaders

Shader-language dialect

Textures

MilkDrop3 extensions

MilkDrop3 (a separate, actively developed fork) extends the format with .milk2 double presets, q33–q64, and shader-side FFT data. None of this is supported anywhere in Stims — confirmed by source search, not just omission. If you're targeting MilkDrop3 specifically, this curriculum and the Stims engine are not the right tool; everything here targets the MilkDrop2-era format MilkDrop3 itself still ships alongside its extensions.

The five things worth remembering

1. Single-pass fusion is Stims's biggest structural divergence — anything depending on a genuinely separate warp-pass buffer gets an approximation.
2. GLSL 1.20, not HLSL — stated plainly, not glossed over, in Track 6.
3. No external textures, period.
4. Sampler aliasing covers roughly 1 in 13 bundled presets — check the coding guide's alias table if a preset references an unfamiliar sampler name.
5. "Supported" here means compiles and renders through Stims's own pipeline — not "measured pixel-identical to a native MilkDrop render." Stims's own compatibility system deliberately keeps those two claims separate; so should you when describing your own preset's portability.

Performance and quality

1. Frame-rate correction is not optional. Any accumulator advanced by a fixed amount per frame (mtime, custom clocks) needs an fps-relative term — x*(75/fps) for anything calibrated against 75fps, or derive your own factor as Track 3's RC filter does (k/fps). Skip it and the same preset runs at different speeds on different machines.
2. Neighbor sampling in a shader costs more than a single sample. Track 6's edge-detection example reads six texture samples per pixel (four neighbors, the pixel itself, and one blurred copy); that's a real, measurable cost multiplier over a one-sample shader, paid by every pixel on screen every frame. Reach for it when the effect needs it — not by default.
3. Fewer per_pixel lines beat more. Every line runs once per mesh point; a preset with a dozen per_pixel equations is doing that work at every vertex, every frame. The coding guide's checklist treats "at least 3 meaningful per_pixel lines" as a floor for depth, not a target to maximize.
4. warp=1 with high decay is the single most common performance-adjacent mistake — not because it's slow, but because it saturates into a smear that makes every other effect illegible. Track 2 covers why.
5. Measure instead of guessing. Track 3, Lesson 5 covers bun run lab:reactivity. Its sibling bun run lab:visual (catalog presets only, currently) adds pixel-level checks (luminance, contrast, colorfulness, and whether the render is blank or static) — the same instrument the Stims project uses on its own bundled catalog.

Publishing

1. Export your preset (or copy its #code= URL — see Track 0).
2. Run it through bun run lab:reactivity -- --file your.milk and aim for a reactive verdict on the variables meant to respond to audio.
3. Open a PR adding the .milk file to public/milkdrop-presets/, or share the link directly — both paths are covered in the contributing guide.
4. If you remixed someone else's work anywhere along the way, keep the credit chain — Remix in the editor's ⋯ menu does this automatically, and the multi-author filename convention ("Author A + Author B — Title") is how the whole community has tracked lineage for over twenty years. It's not paperwork; it's the reason this hobby has a history at all.

What you can now do

Write a preset from the mental model up, wire it to the music, give it depth and shape, reach for a shader when equations aren't enough, know honestly where it will and won't run the same elsewhere, and ship it back into the same catalog you started by browsing in Track 0.

If you want to go further, the essays on palette, pacing, and restraint that turn a technically correct preset into a good one are in Track 7 — Taste.