01 · Introduction

What is Kirlian?

Introducing the first Generative Music Visualizer.

What does it mean to see a song?

Kirlian explores how AI can interpret music and turn that interpretation into a visual experience. It began with my curiosity about constructing systems that could use measurable multimodal inputs in unconventional formats (e.g. music represented through mathematical analysis rather than supplied as audio) to surface properties that might otherwise remain obscure or difficult to express in other mediums.

This particular exploration has opened up a myriad of questions about Generative AI’s “creativity”: How does one consistently measure success in creative, non-deterministic systems? What would a universally successful result even mean? And whose judgment would establish that?

For Kirlian, part of that difficulty lies in the translation from interpretation to aesthetic experience. An evocative description is only the beginning; its ideas still have to take a form that another person can see, hear and feel.

Working on that translation brings together things that don’t usually come in pairs: non-determinism and aesthetics, creativity and data engineering, etc.

To “see” a song is already possible through film, software, code and other artistic mediums. For some people, there is also the experience of chromesthesia.

With AI models, a generative system can work with the context made available to it, make guided decisions about how to interpret that information, and use code to give the interpretation a visual form. Kirlian explores this translation between mediums, with the possibility of producing something that surfaces both the known and the unknown, and can surprise in unexpected ways.

The recording, its context and the listener

Music already carries several kinds of meaning at once. There is the recording itself: its rhythm, instrumentation, texture, structure and development. Then there is the context around it: the artist, the circumstances of its making, its references and associations. Then there is everything a listener brings. The same recording can belong to very different places and periods in different people’s lives.

I wanted to explore what a model could do with those relationships. A personal memory might suggest a material, a colour or a sense of distance. Something in the recording might support that association, complicate it, or suggest another direction. The interesting part is how these things meet.

How interpretation becomes a visual system

Kirlian approaches this through two stages: interpretation and construction.

    • Audio analysisSound, rhythm and structure
    • ContextWhat surrounds the recording
    • Listener notesOptional memories
  1. Stage 1InterpretationWrites the brief
  2. Stage 2ConstructionBuilds, checks and revises the code
  3. Visual systemPlayed in the browser with the song
Fig. 01 From a song to a visual system. Analysis, context and optional listener notes inform a brief; construction turns the brief into a visual system.

Interpretation: writing the brief

The first stage works with a prepared account of the music. Audio analysis supplies information about its structure and changing sound, while contextual material and optional personal annotations provide other ways into the recording. The model uses this material to develop a written interpretation: a creative brief describing a possible visual world and how it should relate to the music.

The distinction between information and interpretation matters here. An audio measurement can describe a change in energy. It cannot decide what that change should look like. That requires a creative choice. The brief makes those choices explicit enough to examine before they become an image.

Construction: turning the brief into code

The second stage turns the interpretation into an executable visual system. The model works with the renderer’s available capabilities to construct the composition and its behaviour, including generated visual code. Construction can involve checking and revising the result. The written brief remains a reference for what the system was trying to express.

Behind one interpretation

Each work comes from one run through a private workflow. The figure follows a single song through it: what runs at the same time, and what has to finish before the next stage can start.

  1. Six at once
    • AudioStems, then analysis
    • Five sourcesFetched together
  2. ContextAssembled by a script, to a token budget
  3. Call 1BriefWritten from the context and notes
  4. Call 2ConstructionBuild, check, observe, revise
  5. SavedOnce a revision is selected
Fig. 02 Behind one interpretation. The six preparation stages run together; the brief waits until all six have finished.

Each step is recorded in the database before it is dispatched. Redis carries the queued work and holds task results; Celery workers execute the tasks, and each lane has its own queue. The order between stages is kept in the database rather than in the queue: a scan waits until all six preparation stages have ended, whether they succeeded or failed, before the context is assembled.

The context document is assembled by a script, not a model, and it is built for token efficiency. It cuts through the noise: from the analysis, the gathered sources and any listener notes, it keeps what the model needs to write the brief, summarises long measurements into a readable shape, and, if the document runs long, drops the lowest-priority material until it fits a fixed token budget. The model starts from the signal rather than the raw data.

Construction is a bounded exchange between the model and its tools. A candidate undergoes technical checks. An accepted revision can be rendered at selected passages, giving the model observations to use in a further revision. An observed revision must be explicitly selected before it can be finalized and saved. Limits or errors can end a run without a finalized result.

All of this is named and implemented to ensure the machine is well oiled for the model, with one honest but important caveat: Passing technical checks does not establish artistic quality.

Records and artifacts are kept at every step: the run and its stages, the stems and analysis, the gathered sources and the assembled context, the brief, each construction revision and its observations, and the published visual system.

Show the full workflow and where it runs
Fig. 02.1The workflow, lane by lane
Each lane is a worker queue. Construction is a subprocess: expand it to see the loop inside.
Fig. 02.2Where the work runs
The workers run as two sets. Audio processing runs on a custom GPU build with an RTX 4090 and a Threadripper.

What happens when you press play

Generation and playback are separate. Once a visual system has been made, the browser executes it alongside the recording, using the song’s timing and audio data. The language model does not generate each frame during playback. It authors the system that produces those frames.

The public demo plays saved copies of these outputs as static files. It needs no account and never calls the private workflow.

    • SavedVisual systemAuthored earlier
    • SavedAudio analysisStructure and stems
    • LivePlaybackThe recording and its time
  1. BrowserRuns the system with the music
  2. FramesNo model call per frame
Fig. 03 When you press play. The browser combines the saved visual system with the analysis and the current playback; the public demo serves saved copies as static files.

Configuration and human judgment

Configuration enters at several points. Choosing what information reaches the model shapes what it has to work with. Personal annotations can introduce an association that would not be available from the recording alone. Visual direction concerns how the interpretation is expressed: through abstraction, a more theatrical treatment, particular colours or materials.

I think it is useful to keep those choices distinct. A memory attached to a song contributes meaning. A choice about visual treatment contributes a way of expressing it. Neither guarantees a compelling result, but both give the work a direction that can be considered and questioned.

My role also extends beyond supplying an initial instruction. I choose what the system is built to support, review the interpretations and visual results, compare versions, and decide what is worth keeping. The model contributes decisions within that process. Part of building Kirlian has been learning where to give it room, where to be precise, and where an apparent creative limitation was actually a limitation I had built into its environment.

How I judge the result

Separating the work into stages gives me somewhere to look when the result is disappointing. Was the interpretation unconvincing? Did construction lose something important? Did the renderer execute what was written? Did it execute correctly but produce something that barely registered on screen? Those are different problems, even when the final impression is simply that the work feels flat.

This has made evaluation part of the creative work. A system can run correctly without producing an experience I want to return to. A written brief can be persuasive while the visual loses its specificity. I need to examine the translation between those stages and also make a judgment as a listener.

Development through a song

One of the things I care most about is development through a song. A visual can respond continuously and still feel unchanged. Brightness, scale and movement may vary while the relationship between its elements stays the same.

I want to explore something more consequential. A returning phrase might bring back an earlier form that has been altered by what happened between its appearances. One element might interrupt another, gradually take its place, or leave something behind. A quiet passage might need very little movement. The aim is to find changes—and moments of stability—that belong to the particular recording.

The first public preview

For the first public preview, I’m curating a small collection of works alongside explanations of selected musical and visual moments. It is an opportunity to encounter the results and understand some of the decisions behind them.

Kirlian is still an experiment in what people and models can make together through music and code. I’m interested in whether a visual interpretation can bring something into focus: a relationship in the music, an association I had not articulated, or a familiar song that becomes unfamiliar enough to hear again.

Published 13 October 2026