NAMFORUM Sysop: Created page

2026-09-01T06:10:10Z

Created page

New page

= Capture Technologies =

'''Capture technology''' is a general term for systems that create digital representations of physical audio equipment by measuring how that equipment responds to known input signals.

Capture systems are commonly used to reproduce:

* Guitar amplifiers
* Bass amplifiers
* Preamps
* Overdrive pedals
* Distortion pedals
* Fuzz pedals
* Compressors
* Speaker/cabinet chains
* Studio equipment
* Complete signal chains

Modern capture technologies include [[Neural Amp Modeler]], [[TONEX]], [[Kemper Profiling]], [[Neural Capture]], [[Line 6 Proxy]], and a variety of [[Open Capture Technologies|open capture technologies]].

Although these systems pursue broadly similar goals, their model formats, training methods, playback engines, hardware requirements, and ecosystems are different.

== Capture versus conventional modeling ==

Traditional algorithmic modeling usually begins with a developer designing a mathematical or computational model intended to reproduce the behavior of a particular device or circuit.

A capture system takes a different approach.

It measures a physical device.

Conceptually:

'''known input → physical device → measured output'''

A training or profiling process then determines a model that reproduces the relationship between the input and output.

The resulting model can process new audio in real time.

== What is being captured? ==

A capture does not necessarily represent one individual piece of equipment.

It represents the behavior of the entire measured signal path.

For example:

'''test signal → amplifier → load box → recorder'''

creates a different measurement from:

'''test signal → amplifier → speaker → microphone → recorder'''

The second path contains the characteristics of the:

* Amplifier
* Speaker
* Cabinet
* Microphone
* Microphone position
* Return signal path

This is why understanding the [[Capture Signal Chain]] is essential.

== Snapshot captures ==

Many capture systems create '''snapshot models'''.

A snapshot represents equipment at a particular configuration.

For example:

'''Amplifier gain: 7'''

'''Bass: 5'''

'''Middle: 6'''

'''Treble: 4'''

'''Master: 3'''

The resulting model represents the measured behavior around that configuration.

Changing playback controls afterward does not necessarily reproduce the exact behavior that would have occurred if the physical amplifier controls had been changed.

For substantially different physical settings, separate captures may be appropriate.

== Conditioned models ==

Some systems extend capture technology by including one or more control values as inputs to the model.

This is known as '''conditioned modeling'''.

Conceptually, a snapshot model learns:

'''audio input → audio output'''

A conditioned model can learn:

'''audio input + control value → audio output'''

This can allow one neural model to reproduce behavior across part of a physical control's range.

Open-source projects such as GuitarML Proteus have experimented with conditioned models.

Commercial systems may also combine captured behavior with separately modeled controls.

== Hybrid capture and modeling ==

The boundary between capture technology and conventional modeling is increasingly blurred.

For example, Kemper's [[Kemper Profiling|Liquid Profiling]] combines a measured PROFILE with modeled gain and tone-stack behavior.

A modern system may therefore contain:

'''capture + algorithmic modeling + conventional effects'''

rather than fitting entirely into one category.

== Major capture platforms ==

Several capture technologies have developed substantial user ecosystems.

=== Neural Amp Modeler ===

[[Neural Amp Modeler]] is an open-source neural modeling system.

NAM provides:

* Open-source training
* Open model specification
* Open-source real-time playback core
* Desktop software
* Third-party hardware implementations
* Community model sharing

NAM models normally use the ''.nam'' file format.

Current NAM snapshot models use the [[NAM A2 Architecture|A2 architecture]], while older A1 models remain supported.

See [[NAM Playback]].

=== TONEX ===

[[TONEX]] is IK Multimedia's proprietary AI Machine Modeling ecosystem.

It provides:

* TONEX Modeler
* TONEX Player
* TONEX Editor
* ToneNET
* TONEX Pedal
* TONEX ONE
* Other TONEX hardware

Captured models are called '''Tone Models'''.

The current TONEX Modeler separates physical capture from model training and can preserve capture data for later processing.

=== Kemper Profiling ===

[[Kemper Profiling]] is the proprietary capture technology used by the KEMPER PROFILER ecosystem.

Captured models are called '''PROFILEs'''.

Kemper also provides:

* Liquid Profiling
* PROFILING 2.0
* Rig Manager
* Rig Exchange
* Dedicated PROFILER hardware

Liquid Profiling combines captured amplifier behavior with modeled control behavior.

=== Neural Capture ===

[[Neural Capture]] is Neural DSP's proprietary capture technology for the Cortex ecosystem.

Current systems support:

* Neural Capture V1
* Neural Capture V2

V1 provides fast local capture.

V2 uses higher-resolution cloud-based training and is intended to improve modeling of difficult dynamic equipment such as fuzzes, compressors, and highly responsive amplifiers.

Neural Captures can be shared through Cortex Cloud.

=== Line 6 Proxy ===

[[Line 6 Proxy]] is Line 6's capture technology for Helix Stadium.

Captured models are called '''Clones'''.

Current Proxy Clone types include:

* Amp+Cab
* Amp
* Preamp
* Distortion

Helix Stadium performs the physical measurement and Line 6's online infrastructure performs model creation.

Proxy Clones can be used in compatible Stadium hardware and software.

=== Open capture technologies ===

Other open-source systems include projects such as:

* AIDA-X
* GuitarML Proteus
* NeuralPi
* SmartGuitarAmp

See [[Open Capture Technologies]].

These projects explore different neural architectures, model formats, training systems, and hardware targets.

They should not automatically be assumed to be compatible with NAM or with one another.

== Comparison ==

The following table provides a broad conceptual comparison.

{| class="wikitable"
! Technology
! Model name
! Ecosystem
! User capture
! Training location
! Sharing
|-
| [[Neural Amp Modeler]]
| NAM model
| Open source
| Yes
| Local / cloud services
| Independent / community services
|-
| [[TONEX]]
| Tone Model
| IK Multimedia
| Yes
| Computer
| ToneNET
|-
| [[Kemper Profiling]]
| PROFILE
| Kemper
| Yes
| PROFILER hardware
| Rig Exchange / independent
|-
| [[Neural Capture]]
| Neural Capture
| Neural DSP Cortex
| Yes
| Hardware for V1 / cloud for V2
| Cortex Cloud
|-
| [[Line 6 Proxy]]
| Clone
| Line 6 Helix Stadium
| Yes
| Line 6 cloud
| CustomTone / file exchange
|-
| [[Open Capture Technologies|AIDA-X / GuitarML / others]]
| Varies
| Open source
| Varies
| Usually computer / cloud workflow
| Varies
|}

This table describes the broad architecture of each ecosystem rather than every possible workflow.

== Model formats are not interchangeable ==

One of the most important facts about capture technology is that there is currently no universal model format.

For example:

'''NAM model ≠ TONEX Tone Model'''

'''TONEX Tone Model ≠ Kemper PROFILE'''

'''Kemper PROFILE ≠ Neural Capture'''

'''Neural Capture ≠ Proxy Clone'''

Even open-source neural models are not necessarily interchangeable.

Each system may use a different:

* Neural architecture
* Model representation
* File format
* Input normalization
* Sample rate
* Metadata
* Playback engine

A model must be interpreted by software that understands its architecture and format.

== Capture versus training ==

The terms '''capture''' and '''training''' are sometimes used as though they mean the same thing.

They are conceptually different.

'''Capture''' is the measurement of the physical equipment.

'''Training''' is the computational process that creates a model from the measurement.

A simplified workflow is:

'''training signal → physical equipment → recorded response → training → model'''

Some systems hide this distinction from the user.

Others make it explicit.

For example:

* NAM training can occur after the physical recording.
* TONEX Modeler can preserve capture data for later training.
* Neural Capture V2 performs physical measurement on Cortex hardware and training in Cortex Cloud.
* Proxy performs physical measurement with Stadium and model creation using Line 6's online infrastructure.

Separating these concepts is useful when discussing capture technology accurately.

== The capture recording ==

The recording produced during the physical measurement is not itself the finished model.

It is evidence of how the equipment responded to the training signal.

Training uses that evidence to construct a model.

Therefore:

'''measurement ≠ model'''

This distinction has important consequences.

If the original measurement is preserved, it may sometimes be possible to train another model from it later.

This can be useful when:

* Training software improves
* A new architecture appears
* A different quality setting is desired
* The original equipment is no longer available

Whether old measurements can actually be reused depends on the requirements of the new training system.

== Training signals ==

Capture systems use carefully designed signals to stimulate the equipment being measured.

A useful training signal must provide enough information for the system to learn the target's behavior.

Different technologies use different training signals and procedures.

They may contain:

* Noise-like material
* Sweeps
* Impulses
* Bursts
* Guitar-like material
* Specially designed excitation sequences

The training signal should not normally be substituted with arbitrary audio unless the particular system explicitly supports it.

== Nonlinear behavior ==

Capture technologies are particularly useful for nonlinear audio equipment.

Examples include:

* Tube amplifiers
* Solid-state amplifiers
* Overdrive
* Distortion
* Fuzz
* Compression
* Saturation

A nonlinear device does not simply apply one fixed frequency response.

Its output depends on the signal entering it.

For example, a guitar amplifier may respond differently to:

* Quiet picking
* Hard picking
* Low guitar volume
* High guitar volume
* Different input levels

A successful capture must reproduce enough of this behavior to remain convincing with new playing material.

== Linear systems ==

Some parts of an audio chain are approximately linear and can be represented efficiently by an impulse response.

The most common example is a:

'''speaker + cabinet + microphone'''

signal path.

A conventional cabinet IR can reproduce its linear frequency and phase response without requiring a neural network.

This is why many capture systems use:

'''neural amplifier model → cabinet IR'''

rather than requiring the neural model to reproduce the cabinet.

See [[Cabinets and IRs]].

== Direct captures ==

A direct amplifier capture excludes the physical speaker and microphone response.

A typical signal path is:

'''training signal → amplifier → suitable load/direct output → recorder'''

The resulting model normally requires cabinet processing when played through a full-range system.

Advantages include:

* Cabinet can be changed later
* IRs can be exchanged
* Physical guitar cabinets can be used
* Amplifier and cabinet can be evaluated separately

See [[Capture Types]].

== Complete rig captures ==

A complete rig capture can include:

* Pedal
* Amplifier
* Cabinet
* Speaker
* Microphone
* Microphone preamp
* Other compatible processing

For example:

'''overdrive → amplifier → cabinet → microphone'''

may be captured as one model.

This can reproduce a specific finished sound very effectively.

The tradeoff is flexibility.

The individual components cannot necessarily be separated after they have been learned as one combined system.

== Capturing pedals ==

Capture technologies can model many gain-based pedals.

Common targets include:

* Boost
* Overdrive
* Distortion
* Fuzz

Pedals can be captured:

* Individually
* With an amplifier
* As part of a larger chain

Capturing them individually provides greater routing flexibility.

Capturing the combined chain can preserve interactions between the devices.

== Difficult devices ==

Not every audio processor is equally easy to capture.

Potentially difficult targets include:

* Fuzz circuits
* Compressors
* Devices with long-term state
* Strongly dynamic circuits
* Devices sensitive to source impedance
* Time-varying effects

A system that performs extremely well on an amplifier should not automatically be assumed to reproduce every compressor, fuzz, delay, or modulation effect equally well.

The architecture and training method matter.

== Time-based effects ==

Snapshot neural capture systems generally focus on relatively short-memory nonlinear systems.

Effects such as:

* Delay
* Reverb
* Chorus
* Flanger
* Phaser
* Tremolo

may require different modeling approaches.

Some capture systems explicitly instruct users not to include these effects during capture.

When capturing a complete amplifier rig, disable time-based onboard effects unless the capture technology specifically supports them.

== Input level ==

Input level is fundamental to capture technology.

For a nonlinear device:

'''changing input level can change device behavior'''

not merely volume.

If the playback system drives a model differently from the original capture conditions, the result may exhibit different:

* Gain
* Distortion
* Compression
* Dynamics
* Sustain
* Cleanup

See [[Gain Staging and Calibration]].

== Calibration ==

Calibration attempts to preserve the relationship between digital signal level and the analog level used during capture.

A calibrated workflow can improve model portability between different audio interfaces and playback systems.

Not every capture platform exposes calibration in the same way.

Some systems manage much of the process internally.

Others provide explicit metadata or require the creator to document the signal chain.

Calibration should therefore be considered part of the capture methodology rather than a universal feature implemented identically by every platform.

== Reamping ==

A capture system often needs to send a recorded training signal from an audio interface into equipment designed for an instrument-level source.

A '''reamp device''' can provide appropriate:

* Level conversion
* Connection
* Ground isolation
* Signal interfacing

The exact requirements depend on the equipment and interface.

See [[Reamping for Capture]].

== Accuracy ==

Capture accuracy is not one simple number.

A model can differ from the original in:

* Frequency response
* Harmonic distortion
* Dynamic response
* Transients
* Compression
* Sustain
* Low-frequency behavior
* High-frequency behavior
* Response to guitar-volume changes

A numerical training error can be useful, but it should not be treated as a complete substitute for listening.

The final model should be evaluated using relevant musical material.

== Level matching comparisons ==

When comparing a capture against the original device, levels should be matched closely.

Human hearing commonly interprets a slightly louder signal as:

* Fuller
* Clearer
* More detailed
* More exciting

An unmatched comparison can therefore produce misleading conclusions.

This principle also applies when comparing two different capture technologies.

== Comparing capture systems ==

A meaningful comparison should control as many variables as possible.

Ideally use:

* Same physical device
* Same physical settings
* Same source recording
* Same cabinet/IR
* Same playback level
* Same monitoring system
* Appropriate calibration
* Equivalent capture type

Otherwise the comparison may measure differences in the surrounding workflow rather than differences in the capture technology itself.

== Model accuracy versus usability ==

The mathematically closest model is not automatically the most useful model.

Practical factors also include:

* CPU usage
* Latency
* Training time
* Hardware cost
* Ease of capture
* Model availability
* Editing
* Preset management
* Live reliability
* Sharing
* Portability

Different technologies make different engineering tradeoffs.

== Native and converted models ==

Some playback systems execute the original model architecture directly.

Others import a model and convert or approximate it using another internal architecture.

Conceptually:

'''native'''

'''model → original-compatible engine → audio'''

versus:

'''conversion'''

'''model → conversion → different model architecture → audio'''

Both approaches can be useful.

They should nevertheless be described accurately.

See [[NAM Playback]].

== Hardware and software ==

Capture models can be played using:

* Desktop plugins
* Standalone applications
* Dedicated pedals
* Multi-effects processors
* Embedded systems
* Mobile/portable hardware

The model and player are separate concepts.

A particular capture technology may support many players or only hardware from one manufacturer.

== Open and proprietary ecosystems ==

Capture technologies span a spectrum from open-source systems to vertically integrated proprietary products.

An open ecosystem may publish:

* Training code
* Playback code
* Model specification
* Development tools

A proprietary ecosystem may integrate:

* Capture
* Training
* Playback
* Hardware
* Cloud services
* Model sharing

Neither structure automatically determines modeling quality.

It does affect:

* Interoperability
* Third-party development
* Long-term access
* Hardware choices
* Model distribution

== Community libraries ==

Capture technology has created large libraries of user-generated models.

Examples include services associated with:

* NAM
* TONEX
* Kemper
* Neural DSP
* Line 6

Community sharing dramatically increases the usefulness of capture systems.

It also creates documentation problems.

A model without information about what was captured can be difficult to evaluate or reuse.

== Model documentation ==

Useful capture documentation includes:

* Platform
* Model architecture/version
* Equipment manufacturer
* Equipment model
* Equipment settings
* Capture type
* Cabinet
* Speaker
* Microphone
* Microphone position
* Load box
* Calibration
* Capture date
* Creator
* Training software/version
* Notes

Not every field applies to every capture.

The objective is to preserve enough information to understand what the model actually represents.

== Model names are not documentation ==

A filename such as:

'''Marshall High Gain'''

does not reveal whether the model contains:

* Amplifier only
* Amplifier and cabinet
* Overdrive and amplifier
* Microphone
* Particular channel
* Particular amplifier settings

Do not infer the capture signal chain solely from a model name.

See [[Capture Types]].

== Reproducibility ==

A well-documented capture can potentially be recreated or investigated later.

Useful records include:

* Original training signal
* Original returned recording
* Equipment settings
* Photographs
* Signal-chain diagram
* Calibration data
* Software versions
* Finished model

Reproducibility becomes particularly valuable when comparing training algorithms or architectures.

== Preservation ==

Capture models may outlive the software or hardware originally used to create them.

For important captures, consider preserving:

* Finished model files
* Original measurements
* Metadata
* Documentation
* Required player versions
* Relevant open-source software where legally appropriate

This is especially important for unique equipment that may later be sold, modified, damaged, or unavailable.

== Capture as documentation of equipment ==

Capture technology has uses beyond replacing physical equipment.

A well-documented model can preserve aspects of:

* Rare amplifiers
* Modified equipment
* Unique pedals
* Studio signal chains
* Artist rigs
* Historical equipment

A capture is not a complete physical simulation of the original device, but it can preserve a useful measurement of how that device behaved at a particular time and configuration.

== Choosing a capture technology ==

The best system depends on the intended use.

Consider:

* Equipment to be captured
* Desired accuracy
* Hardware already owned
* Live requirements
* DAW requirements
* CPU resources
* Training workflow
* Cloud requirements
* Open-source requirements
* Model-sharing community
* Need for conditioned controls
* Cabinet workflow
* Budget

There is no single capture platform that is automatically best for every user or every device.

== NAMFORUM scope ==

NAMFORUM covers capture technology as a field rather than limiting discussion to one platform.

This includes:

* [[Neural Amp Modeler]]
* [[TONEX]]
* [[Kemper Profiling]]
* [[Neural Capture]]
* [[Line 6 Proxy]]
* [[Open Capture Technologies|Open capture systems]]
* Emerging capture technologies
* Capture hardware
* Training
* Calibration
* Reamping
* Model sharing
* Playback
* Research

Conventional algorithmic modeling may also be discussed where it intersects with capture technology, but capture and profiling technologies are the primary focus.

== See also ==

* [[Getting Started with Capture Technology]]
* [[Capture Signal Chain]]
* [[Gain Staging and Calibration]]
* [[Reamping for Capture]]
* [[Creating a Capture]]
* [[Capture Types]]
* [[Cabinets and IRs]]
* [[Troubleshooting Captures]]
* [[Neural Amp Modeler]]
* [[TONEX]]
* [[Kemper Profiling]]
* [[Neural Capture]]
* [[Line 6 Proxy]]
* [[Open Capture Technologies]]
* [[Glossary]]

'''[https://namforum.com/ Discuss capture, profiling, training, and modeling technology on NAMFORUM]'''

[[Category:Capture technology]]
[[Category:Audio technology]]

Capture Technologies - Revision history

NAMFORUM Sysop: Created page