MilkDrop conversion

milkconv turns a MilkDrop .milk file into a Ritmolux preset ahead of time (ADR-0113). This page is how you look at what it produced and decide whether the conversion is good.

It is a developer tool and never ships — it sits outside the workspace’s default-members, exactly as rlx-core-cabi does, so a bare cargo build does not compile it:

cargo run -p milkconv -- "path/to/Some Preset.milk" --out converted.toml

The output is an ordinary preset, so everything on this page already applies to it — shot --preset-file converted.toml renders it like any other:

cargo run -p standalone --example shot -- \
  --preset-file converted.toml --frames 300 --size 640x400 \
  --set bass=0.6,mid=0.5,treb=0.45 --out converted.png

Read the warnings; they are the interesting output

The converter prints every name it could not carry across, to stderr, and repeats them in the bundle’s header comment. That is the plan’s no-silent-zero rule (Phase 3): a preset that reads a variable MilkDrop supplies and this engine does not, or writes one MilkDrop draws with and this engine does not, says so — in the shape ADR-0020’s typo warning already takes.

milkconv: Songflower: custom shape 1 is textured with the previous frame, which this engine has no stage for; it is drawn as a flat gradient fill
milkconv: Cauldron - painterly 5: reads `ang` in per-frame code, where it is a per-VERTEX variable ...

What is not warned about is a bare name the preset uses as its own accumulator. In EEL2 an unset variable legitimately reads zero, presets rely on that constantly, and warning on every one of them would bury the findings that matter. The check is against MilkDrop’s own roster, not against “every name we do not know”.

What a conversion does and does not carry

carried	not carried
the per-frame and per-vertex programs, whole	disk textures — deliberately out of scope, and priced at 19 % of the corpus. Since Phase 6 this is a conversion failure: the shader that samples one is rejected by name rather than silently rendered without it
the warp and comp HLSL blocks, translated to WGSL (Phase 6) — the ~30 intrinsics, swizzles, if, bounded loops, #defines, helper functions, the noise samplers, GetBlur1..3, the q/roam/rand/rot_* input surface	HLSL arrays, computed #if conditions, structs — each a named rejection, together well under 2 % of the corpus
zoom rot cx cy dx dy sx sy warp zoomexp	a custom shape’s textured flag — the previous frame as a fill, which needs a stage this engine has not got
decay gamma wrap darken_center brighten darken solarize invert, and the video echo echo_zoom / echo_alpha / echo_orient (Plan 0109) — a bundle emitted before that plan carries no register for them and must be re-converted	the second audio channel, because this engine’s analysis is mono by construction
the waveform’s eight wave_mode figures and their whole wave_* roster
up to four custom waves and four custom shapes, each with its own programs
the inner and outer borders, and the motion-vector grid
the initial conditions, re-applied at the top of every frame
q1–q32, t1–t8, megabuf, gmegabuf

A converted preset draws its own light and the scene’s deposit stays off. MilkDrop’s light source is the waveform, and up to Phase 3 the converter emitted a stand-in ring in its place so there was something for the mesh to move. That stand-in is gone: the frames further down this section that share a radial-arm character are from that era and are labelled as such.

Rates are converted, and that is why a preset moves at the right speed

MilkDrop’s zoom, rot, dx, dy, warp and decay are all per rendered frame, so a preset drifts at half the speed on a machine running twice the rate. This engine’s vocabulary is per second throughout (ADR-0019). The runtime converts at MilkDrop’s nominal 30 fps — a factor becomes v^30, a rate v * 30 — which is what makes a converted preset move at the speed its author saw, on any display.

Two consequences worth knowing before reading a number in a bundle:

• A zoom of 1.02 in a .milk file is 1.81 per second here. The numbers are not on the same scale as a hand-authored [params] binding beside them.
• fps reports 30 to the program rather than the display’s real rate. A preset that reads fps is compensating for a cadence, and telling it the truth about a 144 Hz display would compensate twice.

The moment of truth, captured

Plan 0100 Phase 3’s done-when is a human judgement — does a converted preset move the way its name says — so the evidence is frames rather than an assertion. All four below are from presets-milkdrop-original, converted with no hand editing and rendered at bass=0.6,mid=0.5,treb=0.45 after 300 frames:

Goody — LSD Zoomtunnel: rays converging on a dark centre, which is the preset’s whole name.	Krash & Rovastar — A Million Miles from Earth (Ripple Mix): the ripple’s radial lobes.
Idiot — Marphets Surreal Dream (Hypnotic Spiral Mix): a spiral, from rot and zoom composed per vertex.	Aderrasi — Contortion (Escher’s Tunnel Mix).

And the same spiral over a clip, which is what “motion in the right direction, at the right rate” actually looks like — six frames of --signal click:120:

The radial-arm character these four share is the stand-in deposit, not the presets. It is the same ring in all of them; what differs is what the mesh does to it. They are kept because they are what Phase 3’s done-when was judged on — read the motion in them, not the figure.

The same thing again with the draw layer (Phase 4)

Phase 4 replaced that stand-in with what MilkDrop actually draws. The figure is now the preset’s own, and the difference is not subtle:

Aderrasi — Contortion (Escher’s Tunnel Mix), the same preset as the last frame of the block above. The tunnel and its dark core are the mesh, as before; the magenta lattice at the edges is the preset’s own custom shapes and waveform, which the stand-in ring had nothing to say about.	Aderrasi — Songflower (Hybrid Plant): four custom shapes and a waveform over a field with fDecay = 1, so nothing ever fades. This is the case that reads as a flat white frame under a purely additive draw seam — see the blend note in Parameter roster.

And with the shaders (Phase 6)

MilkDrop 2’s warp and comp pixel shaders are translated to WGSL at conversion time and run as the preset’s own passes — the warp shader replaces the built-in decay fragment, the comp shader replaces the built-in present. Both frames below are corpus files converted with no hand editing, same settings as above:

Geiss — Blur Mix 3: the waveform stroke sharp, its history softened through the three-level blur chain the comp shader reads with GetBlur1..3 — the look the preset is named for.	Geiss — Myriad Mosaics: the mosaic grid is the warp shader’s cos tiling, the swirl inside each tile its 3D noise-volume lookup (sampler_noisevol_hq), the grain its rand_frame dither.

--report: how much of the corpus converts, and what happens to the rest

Point the converter at a directory instead of a file and it walks it, converts everything, and prints what became of each — Plan 0100 Phase 5:

cargo run -p milkconv --release -- --report path/to/corpus
cargo run -p milkconv --release -- --report path/to/corpus --render   # slower, adds the third count
cargo run -p milkconv --release -- --report path/to/corpus --json     # for diffing two runs

Three counts, and the third is the interesting one:

count	is
parse	the .milk section layout read without error
compile	every EEL2 program in it turned into bytecode
render non-blank	the emitted preset loaded into the engine and put light on screen

Only the third catches a preset that converts perfectly and draws nothing — the failure class the plan’s Risks call the worst for reputation, because it looks like our bug rather than like a refusal. It needs a device, which is why it is behind --render; without an adapter the run reports the other two rather than failing, which is ADR-0016’s policy for captures. Budget about 50 ms a preset.

It asserts no threshold and it exits zero however bad the numbers are. Per ADR-0071 a coverage percentage is a property of one corpus and one converter at one moment, so the report names both and is never a gate. A non-zero exit would make it one the first time somebody put it in a script. The ranked reasons are the output that matters — they are the work list.

The report also prints the two predictions the census made before any of this was built, beside what was measured, so a ranking that disagrees sharply with either is read as evidence about the converter rather than about the corpus. That is the whole reason the census came first, and it costs nothing to state the prediction in advance.

What it said, 2026-08-16

Two readings of the same corpus, one plan phase apart — together they are the record of what Phase 6 changed.

Before the shaders (Phases 1–5). Over all three collections at WORK/milkdrop-corpus, on windows x86_64; 10 325 of 10 347 converted, but the 82 % unconverted-shader row is the honest caveat — four fifths of the corpus rendered with its warp/comp HLSL missing. The full-corpus --render probe (same day) counted 10 257 non-blank of 10 325, with 60 of the 68 blanks declaring MilkDrop 2 — the picture was in the shader nothing ran yet.

  files seen                     10347  100.0 %
  parse                          10347  100.0 %
  compile                        10325   99.8 %
  render non-blank               10257   99.1 %

WHAT A CONVERSION COULD NOT CARRY (top rows):
    8482   82.0 %  unconverted-shader
    2251   21.8 %  disk-texture
                               predicted   measured
  reads a disk texture           19.0 %     21.8 %
  MilkDrop 1.x, no shaders       18.0 %     17.9 %

After the shaders (Phase 6). The translator ran; unconverted-shader is gone, and the disk-texture class moved from a warning to a named conversion failure — a preset whose shader samples a texture we refuse to ship no longer loads without it and renders something its author never drew:

  stage                          count  share
  ---------------------------- -------  -------
  files seen                     10347  100.0 %
  parse                          10347  100.0 %
  compile                         8289   80.1 %
  render non-blank                8063   77.9 %

WHY A FILE DID NOT CONVERT, ranked:
    1217   11.8 %  warp shader disk-texture
     609    5.9 %  comp shader disk-texture
      75    0.7 %  warp shader unsupported
      74    0.7 %  comp shader unsupported
      37    0.4 %  warp shader parse
      16    0.2 %  comp shader unknown-name
      15    0.1 %  per_frame
       6    0.1 %  warp shader unknown-name
       5    0.0 %  per_vertex
       2    0.0 %  comp shader parse
       2    0.0 %  per_frame_init

Read together: 8 289 presets (80.1 %) now convert whole, shaders included, and 8 063 (77.9 %) render non-blank under the probe — the converted set is about 97 % of everything not bound to a disk texture. The unsupported and parse rows (under 2 % combined) are HLSL arrays, computed #if conditions and similar exotica, each rejected by name. Of the 226 converted-but-blank presets, 218 declare MilkDrop 2 — their shaders ran, so those are shader-fidelity findings — and the 8 shaderless blanks are the same 8 the pre-Phase-6 probe found. On the 552-preset original pack, where disk textures are rarer, the same run reads 476 converted (86.2 %) and 471 non-blank (85.3 %).

Both predictions held, which is the result that says the ranking is about the corpus. The shaderless share landed at 17.9 % against a predicted 18 % — as close as a prior gets. The disk-texture class came in 2.8 points above its prediction, and that direction is expected rather than surprising: the census counted a grep for a texture name, and the converter also flags sampler_pc, so it sees a slightly wider class than the census could.

“Converts” is still a weaker claim than “renders as authored” — 82 % of the corpus carries HLSL that nothing translates yet, which is the number Phase 6 is sized by.

The eleven files that do not compile are genuinely malformed — a stray backslash, a sin with no arguments, a truncated per_frame_init. Two decisions in the parser are what got the number that high, and both were measured rather than assumed; they are on milk::join_code and eel::Compiler::argument.