Why Static Captures Are Not Enough

The neural amp modeling community has produced over 90,000 capture files. Not one of them has a working knob.

THE SNAPSHOT PROBLEM

Neural Amp Modeler (NAM) is an open-source project that trains neural networks to replicate the input-output behavior of guitar amplifiers. You run a calibrated signal through your amp, record the output, and train a model. The result is a VST plugin that sounds like your amp — at that exact setting.

Want the treble at 7 instead of 5? Run the capture again. New model. Want the gain at 3? Again. Each model is a frozen snapshot of one specific configuration. TONE3000, the community hub, hosts over 90,000 of these files.

Neural DSP takes a different approach — they build amp simulators with working controls — but each amp is a separate $99 to $199 purchase. Static circuit modeling. The product is a collection of individual, isolated amp plugins.

PARAMETRIC VS NON-PARAMETRIC

The distinction matters. A non-parametric model learns a fixed mapping: input signal → output signal. It cannot generalize beyond the exact conditions under which it was trained. Change a knob, and you need a new model.

A parametric model learns a conditional mapping: (input signal, parameters) → output signal. The parameters — knob positions — are inputs to the model, not fixed constants. The model learns how the amp responds to parameter changes, not just how it sounds at one setting.

This is the difference between a photograph and a living, breathing person. A photograph captures one moment. A twin captures the entire behavior.

THE PANAMA FOUNDATION

The academic groundwork for parametric neural amp modeling was laid by the PANAMA paper (Parametric Neural Amp Modeling Architecture). It demonstrated that conditioning a WaveNet-style architecture on continuous control parameters could accurately reproduce amplifier behavior across the full parameter space.

ToneKeep builds on this foundation. Our Pi-NAM (Physics-Informed Neural Amp Model) architecture uses Feature-wise Linear Modulation (FiLM) for parameter conditioning, Snake activation functions for tube saturation modeling, and dilated causal convolutions with a 6,139-sample receptive field to capture the long-range temporal dependencies in tube amplifier circuits — sag, compression, and bloom.

WHAT THIS MEANS FOR YOU

You install one plugin. You get the entire amplifier. Treble, Bass, Volume, Mid-Bite — every knob works, continuously, in real-time. No presets. No menu-diving. No re-capturing. You turn the virtual knob and the physics responds exactly as the real circuit would.

393 KB. Zero latency. Every knob works.

Not a snapshot. A twin.