valve
AI-assisted project. This codebase was created with Claude
(Anthropic), directed and reviewed by a human author. The valves are solved rather
than drawn, and the solution is measured: an offline harness drives the real plugin
class in a headless GL context and reads each claim back out of the picture it made,
from each valve's gain at its operating point and the grid-current knee to the
harmonics, the crossover notch and what happens to hue. Eight negative controls and
six one-character mutants show the checks can fail. It has never been loaded into
Resolume on macOS; on Windows it passed all nine of the fleet Arena gate's checks.
On macOS it's loaded by oxbow, which is a
proper FFGL host but isn't Resolume. The numbers are under Status.
Your picture through a valve amplifier, run like a guitar through an overdriven amp, as an
FFGL effect for Resolume Arena and Avenue. Turn the drive up and it
crunches the picture the way an amp crunches a guitar: shadows crushed, highlights hot,
colour saturated.
One frame of Resolume's bundled clip IntoTheGlow_02, rendered by vatest,
the offline harness, rather than captured from Resolume. Left: the clip. Right: the
defaults, which are two 12AX7 stages into a push-pull pair of EL34s, 1.2 V of drive,
resting at 0.3.
Download
v0.1.0 — prebuilt for macOS and Windows. Pick your platform:
All builds, checksums and release notes: github.com/stoatworks-labs/valve/releases.
macOS builds are signed and notarised and open normally. The Windows builds are unsigned, so SmartScreen warns once.
The one idea
Nobody draws a valve's transfer curve. It's where the valve's plate-current law
meets the circuit's load line. Put the video level on the grid of a real stage
(a valve, a plate resistor, a supply, the next grid's leak) and solve for the plate
voltage. Then chain the stages the way a guitar amp does: preamp triodes, the loss
of a tone stack between them, a phase inverter, and a power pair into an output
transformer.
The valves are Norman Koren's SPICE models (1996), with his parameters:
the 12AX7, 12AT7, 12AU7 and 6DJ8 for the preamp; the EL34, 6L6GC, KT88
and the 300B for the power stage.
What falls out
None of this is drawn on:
- Gain is the operating point's. A 12AX7 stage has a gain of 59 and
clips at a volt; a 12AU7 in the same socket has 13 and stays clean. That's
the swap guitarists make, and it works here for the same reason.
- The two sides clip differently. Toward cutoff the current fades
softly. Past 0 V the grid conducts, loads whatever is driving it, and the
stage's slope drops by r_g/(r_g + R_s), 35 times, at a knee.
- Every stage inverts, so the asymmetry alternates down a chain, which
is why two stages sound (look) different from one. As Wired shows the
inversions: one stage is a negative.
- One valve makes even harmonics; a pair cancels them. A single-ended
stage's distortion is mostly second harmonic. A push-pull pair is odd about
its rest, so the second harmonic vanishes and the third stays; mismatch the
pair and it comes back, in proportion.
- Bias a pair cold and you get crossover. At 5 % idle both valves are
nearly off at rest: the centre of the curve has a third of the slope of
its neighbours, a flat step at mid-grey. Bias it hotter and it fills in.
- Y/C keeps the hue. Chroma travels as a subcarrier. Through any valve
its amplitude compresses (colour saturates, then limits), but its phase,
which is the hue, can't move. Put RGB through the same valves and every
channel clips separately, so hues swing toward the primaries.
- Composite is differential gain. Luma and the subcarrier share one
valve, so the colour's gain is the curve's slope at the brightness it rides
on, and the curvature rectifies part of a strong colour into its own luma.
Six settings on the harness's card, with Show Curve on: the white curve is
the picture's transfer, the magenta one (Y/C) the chroma's amplitude, and in
RGB a mismatch draws three. Rendered by vatest, with the labels added
afterwards.
Watch it (58 seconds):
Drive from clean to fuzz, the 12AX7-for-12AU7 swap, one preamp stage to four,
Bias from cold to hot, a push-pull pair biased into crossover, RGB against Y/C on
the same skulls, Mismatch splitting the curve in three, and Rest Level rescuing a
dark clip, with Show Curve drawing the transfer in the corner throughout. It's all
the plugin's own output. An FFGL plugin doesn't have a window to film, so the footage
comes from this repo's offline harness (vatest --pipe, driven by a cue sheet) rather
than a screen, and the clips are Resolume's bundled demo media.
Controls
| Group | |
|---|
| Signal | Signal (RGB / Y/C / Composite), Rest Level (the picture level that sits at the valves' rest — where a black-level clamp would hold it; 0.3 by default, lower for dark footage), Interstage (−40 to 0 dB between preamp stages: the tone stack's loss), Mismatch (0–25 % spread between valves: R, G and B become three valve sets, and each pair's halves differ) |
| Luma / RGB | Stages (0–4 preamp triodes), Preamp (12AX7, 12AT7, 12AU7, 6DJ8), Drive (volts on the first grid for black to white, 10 mV to 316 V), Bias (the preamp's rest from cutoff to 0 V), Power Stage (Off, Single-ended, Push-pull), Power Valve (EL34, 6L6GC, KT88, 300B), Master (the phase inverter and master pot: volts on the power grids per volt of preamp swing, 0.01 to 100), Power Bias (idle dissipation, 5 % to 100 % of the valve's rating; 70 % is where a tech sets it) |
| Chroma | The same eight, for the chroma's own amplifier in Y/C |
| Output | Output (Fit: black stays black and white stays white; Unity: the operating point's small-signal gain is 1, so drive compresses; Plate: the plate's own volts over its supply), Polarity (Corrected / As Wired), Show Curve, Mix |
Start with Drive, then Stages and Preamp. Turn Rest Level down
when a clip is dark and everything goes to black: that's the shadows
sitting in the valve's cutoff.
Status
v0.1.0, 4 October 2026, and still early days.
Measured offline, on macOS
tools/verify.sh passes on this Mac (M4 Max, macOS 26.4.1) against a fresh
universal Release build. It runs every rendered check at two rasters,
320×180 and 1280×720, and again on Apple's software renderer at 320×180.
The numbers below are at 1280×720.
| check | result |
|---|
--gain | the picture's slope at mid-grey in Plate mode against small-signal theory from Koren's own partial derivatives: 12AX7 −58.900 (theory −58.901), 12AT7 −39.845 (−39.847), 12AU7 −13.0542 (−13.0542), 6DJ8 −23.015 (−23.015); EL34 single-ended −57.776 (−57.776), push-pull 8.7404 (8.7405); 6L6GC, KT88 and 300B the same to four or five figures |
--knee | across the grid-current knee the slope drops to 0.02852 of itself (12AX7) and 0.02847 (12AU7) against r_g/(r_g + R_s) = 0.02857; the sharpest bend in the ramp is at the column where the grid crosses 0 V |
--harmonics | one 12AX7: H2/H1 0.01702 against Taylor's 0.01697, H3/H1 0.000517 against 0.000514; an EL34 pair: H2/H1 4e-8 matched (the float's floor is 6e-6) with H3/H1 0.031; 2 % mismatch gives H2/H1 0.00113, 4 % gives 0.00227 (×2.002) |
--crossover | at five idle settings the centre slope is the composite load line's (worst 0.03 of tolerance); the notch's depth — centre slope over its neighbours' — is the model's, 0.336 at 5 % idle rising to 0.896 at 100 % |
--hue | Y/C: 96 patches, hue kept to the float's bound; the chroma's amplitude is the describing function (4096-point quadrature of the model) within 0.62 of tolerance, and falls as amplitude rises; RGB through the same valve turns hue by up to 0.58 rad |
--dg | Composite: the chroma's gain up a seven-step modulated staircase is f′(Y)/f′(0.5) (worst 0.01 of tolerance), a differential gain of 48 %; a 0.2 carrier lifts its luma by +0.00911 against f″a²/4 = +0.00905; in Y/C the gain does not move with luma (6.6e-6) |
--polarity | as wired, ten chains have the sign their inversions give; corrected, all rise; an inverting chroma chain turns every hue by π to 3e-7 rad |
--identity | no valves is a wire in all three modes (worst 6e-7); 10 mV on a 12AU7 in Unity is the input to 1.7e-4, inside the curve's own departure; Mix 0 is the input bit for bit; premultiplied alpha is kept |
--reference | every pixel of six settings against the CPU running the plugin's own tables through the shader's arithmetic, worst 0.16 of a bound walked lookup by lookup; and the tables against the harness's own valves (its own Koren, operating points and solves), worst 0.77 of the interpolation bound |
--curve | Show Curve draws exactly the curve the picture follows (0 px) and changes nothing outside its box |
Render cost (GPU timer queries, median of 60 frames, on a GPU shared with
other builds, so the figures move by up to 2× between runs): 0.1–0.2 ms at
720p and 1080p and 0.55 ms at 4K for the defaults; Y/C and Composite add a
small describing-function pass, 0.9 ms at 4K at most. Solving the tables
again when a valve, bias or mismatch setting moves costs about 10 ms on the
CPU for a whole chain (only the stages a setting touches are re-solved);
Drive, Master, Interstage, Stages and Rest Level cost nothing. These are macOS
figures only.
Not established
It has never been loaded into Resolume on macOS. Everything above was
compiled, rendered and measured offline against the real plugin class in a
headless CGL context, plus a load in oxbow. On Windows, a build of this source
passed all 9 of the fleet Arena gate's checks on win-lab (Resolume Arena 7.27.1,
Mesa llvmpipe, no GPU, 2026-10-04): it loads from Extra Effects, registers as
SW Valve / VA01 / effect, all 30 host controls match the declaration (the two
names at exactly 16 characters arrive whole), it renders, every control (and Resolume's own
Opacity: 25) moves the picture, and Arena's log stays clean. Software rendering says
nothing about a GPU or about speed. I haven't tested how it looks on real footage at a
show, how the controls read in Arena's inspector on macOS, or whether a 10 ms table solve
is noticeable while you drag Bias. Koren's models are fits to published curves, good to a
few per cent, and the circuit values are typical guitar-amp ones rather than any one
amplifier's. The model is memoryless: the coupling capacitors' blocking, the
Miller capacitance that would smear a 12AX7's output horizontally (and take a
4.43 MHz chroma carrier out entirely), supply sag and the output transformer
aren't in it. There's a user guide.
No OpenFX port, as it wasn't in scope for 0.1.0. There is a
browser demo now.
Browser demo
https://valve-demo.stoatworks-labs.com/ — every control with its own name, group and
default. The describing-function and display passes are the plugin's own GLSL, spliced
unedited into demo/shaders.js (demo/tools/check_shaders.py, run by tools/verify.sh,
fails on a changed character). The CPU half is a JavaScript port: Koren's valves, the load
lines and operating points (Valves.cpp), the warped tables and the normalisation
(Chain.cpp), the control laws, and the frame's uniforms. demo/tools/crosscheck/ drives
the real plugin class in a GL context through 102 settings and requires the port to
print the same tables, operating points, normalisation and uniforms (read back out of the
plugin's programs), to the bit once the plugin's libm results are replayed into it; on the
browser's own Math every float the GPU reads is still identical. Thirteen frames compared
with vatest on the same input came out pixel for pixel the same but for one alpha-edge
pixel. It needs WebGL2 with EXT_color_buffer_float and a 4,097-texel-wide texture. Stages
is a dropdown of its five values. It's a demo, not the plugin, and the page says what it
doesn't reproduce.
Build
You'll need CMake 3.15+, a C++17 compiler and the FFGL SDK submodule.
git clone --recursive https://github.com/stoatworks-labs/valve
cd valve
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --parallel
cmake --install build # into ~/Documents/Resolume Arena/Extra Effects
macOS builds are universal (Apple Silicon + Intel) by default; add
-DCMAKE_OSX_ARCHITECTURES=arm64 for a quicker dev build. Windows needs GLEW via vcpkg.
Building and testing
The offline harness renders the real plugin class headlessly:
./build/vatest --out /tmp/frame.png --size 1920x1080 # the test card
./build/vatest --list # every control, kind and default
./build/vatest --gain --knee --harmonics --crossover # each claim, measured
./build/vatest --hue --dg --polarity --identity --reference --curve
./build/vatest --negative # and the checks can fail
./build/vatest --offline # what needs no GL (CI)
./build/vatest --bench # 720p, 1080p and 4K
python3 tools/sweep.py # no control is silently dead
tools/mutate.sh # one character at a time
tools/verify.sh # all of it, on a fresh universal build
License
MIT. See LICENSE and ATTRIBUTIONS.md.
This project is built on other people's work — see ATTRIBUTIONS.md.
