INITIALIZING SYSTEM_
X: 0000 Y: 0000
FPS: 60
SCROLL: 0%
SOUTH HUSTLES · CUSA · 2026
STATUS: ONLINE_
Three.js Scope H100 TouchDesigner Resolume NDI ControlNet LoRA
Scroll Progress
0%
Active Section
HOME
Pipeline Load
108GB
Inference
H100 REMOTE
Daydream Cohort 3 · Interactive AI Video Program

PERRO
ANDALUZ

Real-Time AI Visual System for Live Rock

_
Enter the System > View on Daydream >
Scroll
TOUCHDESIGNERSCOPE H100RESOLUME ARENANDI PROTOCOL108GB PROCESSEDZERO LOOPSPERRO ANDALUZCÓRDOBA ARGENTINA TOUCHDESIGNERSCOPE H100RESOLUME ARENANDI PROTOCOL108GB PROCESSEDZERO LOOPSPERRO ANDALUZCÓRDOBA ARGENTINA
01
01 — The Eye

Vision is not a simple act of looking.
It is a dialogue between the eye, the mind,
and what is being looked at.

The central symbol of this project is not a logo. It is an organ. A system that receives vibrations and constructs experience — like the retina constructs color from electromagnetic waves.

"Todo comienza con la luz. Antes de que hubiera palabras, antes incluso de que hubiera dioses, el ser humano abrió los ojos y descubrió que el mundo no era gris — que estaba hecho de tonos que ardían, que se deslizaban, que respiraban."

"El color es lenguaje emocional, un idioma que no necesita gramática porque se dirige directamente al cuerpo."

Civilización Mental — On Color, Light & Perception

Why build an eye?

Newton proved white light fractures into spectrum. Neuroscience proved the spectrum doesn't exist outside your skull — the retina receives electromagnetic waves and the cerebral cortex invents the color you experience.

Our pipeline mirrors this exactly. The audio signal enters TouchDesigner as raw vibration. Scope's inference engine on H100 GPUs constructs a visual reality that never existed before that millisecond. Resolume composes it into shared experience — projected light hitting 500 pairs of eyes.

The pipeline doesn't generate visuals. It generates perception. Every frame the audience sees was born in that exact second. It will never exist again.

See the Pipeline >
The Retina

TouchDesigner

FFT spectral analysis dissects the live audio. RMS detects dynamics. Onset detection triggers prompt sequences on every kick drum. The sensory organ of the system.

[ TD INTERFACE ]
The Cortex

Scope · H100 Remote Inference

LongLive V2V pipeline. Custom LoRAs per track. The model doesn't reproduce — it constructs. Frame by frame, a new visual language.

[ SCOPE UI ]
The Experience

Resolume Arena

4K masks. Layer composition. Post-process. Three.js geometry as visual anchor. The moment raw inference becomes projected light.

[ RESOLUME COMP ]

"El color no es una pared que se interpone, sino un puente. A través de él aprendemos a reconocer nuestra fragilidad, nuestros deseos, nuestra mortalidad."

"El color es fugaz, cambia con la luz, desaparece con la sombra — pero en ese instante de resplandor sentimos que la vida se intensifica, que el mundo nos habla directamente."

The bridge between sound and image is the same bridge between light and perception.
NO DECORAMOS SHOWSLOS PROGRAMAMOSCADA TEMA TIENE SU PROPIA PIELEL CAOS TIENE ESTRUCTURALA ESTRUCTURA TIENE ALMA108 GIGAS SIN REPETIRSE NO DECORAMOS SHOWSLOS PROGRAMAMOSCADA TEMA TIENE SU PROPIA PIELEL CAOS TIENE ESTRUCTURALA ESTRUCTURA TIENE ALMA108 GIGAS SIN REPETIRSE
System Overview

The retina receives.
The cortex constructs.
The experience projects.

Three organs. One nervous system. The pipeline mirrors how human vision works — from raw electromagnetic input to constructed perception. No step is automated. Every frame passes through human hands.

Input

Live Audio

4ch

Analysis

FFT Bands

6

Generation

Inference

H100

Output

Resolution

4K

Pipeline Status
LIVE — Córdoba, Argentina
02
02 — The Pipeline

The Anatomy of
Real-Time Hallucination

The audio signal enters. The system listens, dissects, transforms, composes. No loops. No templates. The output is alive because the input is alive.

"Artistas como James Turrell dejaron de pintar con pigmentos para trabajar directamente con la luz, sumergiéndonos en atmósferas que transforman la percepción."

From Turrell's light rooms to H100 inference — working directly with the raw material
01 · Input
Audio Signal
Live feed — guitar, drums, bass, voice. Unprocessed. The raw material driving every pixel.
02 · Analysis
TouchDesigner
FFT by band. RMS dynamics. Onset on kick. Prompt sequencing "por tirada" — 100+ terms. Output via NDI.
03 · Generation
Scope · H100
Remote Inference. LongLive V2V. Custom LoRAs per track. Denoise + structure live via UI sliders.
04 · Output
Resolume Arena
4K masks. Color grade. Three.js anchors. Composition per scene. LED / projector / OBS.
>>>
01input: "live_audio" // guitar + drums + bass + voice
02analysis: TouchDesigner { fft: "by_band", onset: "kick", rms: true }
03prompts: 100+ terms, sequenced "por_tirada" // per song section
04transport: NDI // zero capture cards, pure network
05inference: Scope.H100 { model: "LongLive_V2V", lora: "perro_andaluz_v3" }
06control: { denoise: slider, structure: slider, canny: slider, depth: slider }
07output: Resolume { scenes: 23, masks: "4000x2000", target: "LED+projector" }
08result: 108GB // zero repeated frames. zero loops. pure instinct.

The system is an instrument

There's a phrase from my father: "If you have ten hours to cut down a tree, spend nine sharpening the axe." We invested weeks crafting 18 structured prompts, training LoRAs, designing 4K masks, hand-coding Three.js geometry — before the first live frame was generated.

Moving UI sliders during a live show is performing an instrument. You shift denoise strength and the hallucination breathes differently. ControlNet — Canny reads the geometry of bodies on stage, Depth maps force the algorithm to respect physical space.

The 12-frame latency becomes a musical breath. The system doesn't react instantly — it inhales with the band and exhales light.

  • Audio-reactive prompt sequencing — the kick drum is the trigger
  • 100+ term prompt table organized "por tirada" per song section
  • NDI protocol — pure network, zero capture cards
  • ControlNet

    ControlNet

    Canny edge detection reads silhouettes. Depth maps force the AI to respect 3D space. Together they prevent the hallucination from overriding human geometry on stage.

    : Canny + Depth preserving human geometry
  • Custom LoRAs

    LoRA Training

    Low-Rank Adaptation models trained specifically on Perro Andaluz's aesthetic. The AI speaks the band's visual language — underground grit, identidad de barro.

    — the AI speaks Perro Andaluz
  • Three.js Iris + lateral fluid ovals as geometric anchors
Technical Architecture

TouchDesigner configured as a sensory organ. Sub-bass drives fluid simulations, mid-range modulates prompt intensity, high-frequency triggers particle density. The FFT feeds a dynamic prompt sequencer indexing 100+ visual descriptors.

Signal pushes via NDI to Scope on H100 clusters running LongLive. The interface: denoise strength, adherence weight, ControlNet influence, LoRA blend — all manipulated live by hand. No automation. Human touch driving cloud inference.

Resolume composites with 23 handcrafted scenes. Each carries its own mask topology — 4000x2000px PSDs preventing the latent space from overwhelming the band's image. Three.js provides structure: central Iris flanked by lateral ovals that transform with the feed.

0GB
Processed
0
Repeated frames
0
Live scenes

el lienzo no es una pantalla blanca, es el audio crudo mordiendo el cable. haciendo real-time parameter shift al milisegundo. tocas la viola y el sistema sangra pixeles. 108 gigas procesados end-to-end y ni un frame se repite.

FFT SPECTRAL ANALYSISRMS DYNAMICSONSET DETECTIONPROMPT SEQUENCING POR TIRADAH100 TENSOR CORESLONGLIVE V2VCANNY EDGEDEPTH MAPCUSTOM LORA4000x2000 MASKSNDI PROTOCOLZERO LATENCY CAPTURE FFT SPECTRAL ANALYSISRMS DYNAMICSONSET DETECTIONPROMPT SEQUENCING POR TIRADAH100 TENSOR CORESLONGLIVE V2VCANNY EDGEDEPTH MAPCUSTOM LORA4000x2000 MASKSNDI PROTOCOLZERO LATENCY CAPTURE
03
03 — The Show

Each Track Has Its Own Skin

Perro Andaluz — 15 years of experimental rock from Córdoba, Argentina. The sound always demanded a body. Now we gave it a nervous system.

"El rojo acelera el pulso, el azul calma la respiración, el amarillo despierta la mente. Jung los veía como arquetipos — símbolos primordiales que emergen del inconsciente colectivo."

"Kandinski escribió que el color es la tecla, el ojo es el martillo, y el alma un piano de cuerdas infinitas."

Every track is a different archetype. Every song, a different key on the infinite piano.

Córdoba: The Interior

While the coasts look outward, the center of Argentina looks inward. 15 years distilling a sound that needed its visual equivalent. Artistic direction by Paulina Olivero — co-director of Relatos Salvajes. "Treat every kick drum like a jump cut. Treat every song like a scene."

The system was born in Patagonia (Cohort 2 — processing nature from a solar-powered motorhome) and evolved into the electric chaos of a live rock stage. Same pipeline, different input. That portability is the proof.

Cohort 2: Patagonia>Cohort 3: Rock Stage First we textured the mountains. Then we textured the distortion.
Cohort Evolution > Paulina Olivero >

Visual identity per track

Heavy Distortion

Aggressive Pulse

Red palette. Maximum prompt rate. Every kick triggers full frame mutation. LoRA trained on underground grit — identidad de barro.

Atmospheric

Sustained Hallucination

Blue-to-purple gradient. Slower denoise decay. Guitar sustain stretches the visual — the hallucination lingers as long as the note resonates.

Cinematic

Scene Tension

Deep blacks, cyan accents. Paulina's direction — every silence is a long take, every explosion is a cut. ControlNet locks the singer's silhouette.

Natural Fluids

Winamp Memory

Remember staring at Winamp as a kid? TD fluid simulations are the purest audio-to-image truth. We rebuilt that purity on cloud GPUs.

te acordás mirando el crt de pendejo re hipnotizado con el winamp? construimos exactamente ese state of mind pero con un tech stack brutal. music played with our bare eyes.

DONDE EL SONIDO TOCA LA MATERIAWHERE NATURE AND TECHNOLOGY COLLIDECODING MY OWN REALITY AS AN INPUTVISUALS THAT NEVER REPEAT DONDE EL SONIDO TOCA LA MATERIAWHERE NATURE AND TECHNOLOGY COLLIDECODING MY OWN REALITY AS AN INPUTVISUALS THAT NEVER REPEAT
South Hustles × CUSA

Built by hand.
Performed by hand.
Projected as light.

Franco Altavista. Córdoba, Argentina. The same hands that code the Three.js, train the LoRAs, and design the masks are the hands on the sliders during the live show. No delegation. Pure tracción a sangre.

C2

Patagonia

Solar-powered motorhome. Nature as input. The system's first breath under open sky.

C3

Rock Stage

Live band. 108GB. Same pipeline, different reality. The proof of portability.

WEB

Three.js

Hand-coded Iris, lateral ovals, particle systems. The stage translated to browser.

"Cada vez que desaparece un lenguaje, desaparecen cientos de colores que solo esa cultura podía nombrar. El color no es universal — es tan local como el acento, tan íntimo como la piel."

From global infrastructure to local identity — del interior profundo al latent space
SOUTH HUSTLESCUSACÓRDOBA ARGENTINADAYDREAM COHORT 3PERRO ANDALUZPAULINA OLIVERO15 YEARS108GB ZERO LOOPS SOUTH HUSTLESCUSACÓRDOBA ARGENTINADAYDREAM COHORT 3PERRO ANDALUZPAULINA OLIVERO15 YEARS108GB ZERO LOOPS
04
04 — The Manifesto

Don't Decorate the Stage.
Program the Organism.

A static video loop is an insult to live music. This system was built to kill the loop — every frame as unrepeatable as the moment it was born in.

"Rothko llenó salas enteras de campos de color que no se miran — se habitan. Un rojo profundo podía hacerte llorar sin que supieras por qué."

"El color es la voz más antigua y más íntima del universo — el alma secreta del mundo."

The secret soul of the world — Manifesto

The shift

99% of live concert visuals are dead on arrival. Pre-rendered stock. Loops repeating every night. VJs mixing clips without listening. The show sounds live. The visuals don't.

We eliminated the local GPU. Scope's Remote Inference on H100 clusters — the same infrastructure that trains foundation models, not a gaming laptop backstage. Processing power in the cloud. Creative control in the hands.

The web experience mirrors the stage. Three.js Iris focuses attention. Lateral ovals shift between organic and electric. The architecture breathes in sync — the 12-frame gap is not a limitation, it's a rhythmic feature.

What we built

  • TD > NDI > Scope H100 > Resolume > LED/Projector
  • 18 manually crafted prompt structures — creative direction as code
  • Custom LoRAs per track — the model speaks the band's language
  • 4000x2000px layered masks — handcrafted precision
  • Three.js Iris + lateral fluid geometry
  • ControlNet: Canny + Depth preserving silhouettes
  • Motion capture — style transfer from real movement
  • 108GB — zero loops, zero repeated frames

What it proves

  • Live visuals fully generated in real-time — no pre-rendered content
  • Same pipeline works across radically different inputs
  • UI slider control is a performance instrument
  • Cloud inference eliminates local GPU on stage
  • Handcrafted direction prevents generic output
  • 12-frame latency is a rhythmic feature, not a bug
The bet

If the audio signal is rich enough and the creative direction precise enough, cloud inference produces visuals more alive than any library. Not because the AI is smarter than a VJ — but because the VJ is inside the loop, hands on sliders, performing the system while the band plays.

The human touch is amplified across H100 tensor cores. Every slider movement propagates through inference and returns as light. The audience doesn't know what's generated and what's real. That's the point.

0
Prompts crafted
0
Years of rock
0
px mask width
View on Daydream > Back to Top >

cortemos con la gilada del video mp4 en loop atrás del cantante. no decoramos más, programamos. el input es el show de rock entero. the output is pure handcrafted hallucination.

Community

Daydream Cohort 3 — the people who build in public.

Franco Altavista — South Hustles / CUSA · Pipeline, Creative Direction, Web

Paulina Olivero — Artistic Direction · Co-director Relatos Salvajes

Perro Andaluz — 15 years of rock, Córdoba, Argentina

Scope TouchDesigner Resolume Three.js NDI ControlNet LoRA LongLive
// TouchDesigner

TouchDesigner — The Retina

FFT spectral analysis dissects the live audio into frequency bands. Sub-bass (20-80Hz) drives fluid simulations. Mid-range (200-2kHz) modulates prompt selection intensity. High-frequency content (4kHz+) triggers particle density changes.

Every kick drum onset acts as a trigger for prompt sequencing "por tirada" — indexed against a curated table of 100+ visual descriptors. The signal pushes via NDI (Network Device Interface) — zero capture cards, pure network delivery at minimal latency.

TouchDesigner doesn't just "listen" to the audio. It dissects it. The result is structured data that drives every downstream decision in the pipeline.

FFTRMSNDIOnset
6
Freq bands
100+
Prompt terms
<5ms
NDI latency
// Scope · H100

Scope — Remote Inference

Scope Desktop connects to H100 GPU clusters running LongLive / StreamDiffusion V2 pipelines. Video-to-video transformation happens in real-time — no local NVIDIA GPU needed on stage.

The critical interface is the UI: denoise strength controls how much the model "hallucinates" vs respects the input. Structural weight controls ControlNet adherence. LoRA blend ratios shift the aesthetic per track. All manipulated live by hand during performance.

Custom LoRAs trained specifically for Perro Andaluz ensure the output doesn't look like generic AI art. It looks like the band's visual language — trained on their grit, their aesthetic, their underground identity.

H100LongLiveLoRA
H100
GPU cluster
12
Frame latency
3
Custom LoRAs
// Resolume Arena

Resolume — Composition

Resolume Arena receives the generated video stream and composites it with 23 handcrafted scenes. Each scene carries its own mask topology — 4000x2000px layered PSDs designed to prevent the latent space from overwhelming the band's real image.

Post-processing includes color grading per scene, blur and blend modes, and real-time mix between the AI-generated feed and natural TouchDesigner fluid simulations.

The Three.js framework provides geometric structure on the web side: a central hand-coded Iris flanked by two lateral conical ovals (OVALADOS) that physically transform alongside the video feed.

4K Masks23 ScenesThree.js
23
Scenes
4K
Mask res
3
Output targets
// Audio Signal

The Raw Input

Guitar, drums, bass, voice — four instruments creating uncompressed live audio. No backing tracks. No click track. Pure analog signal hitting the digital pipeline in real-time.

The unpredictability is the feature. A drummer rushing the tempo means the visuals accelerate. A sustained guitar note means the hallucination lingers. The system doesn't follow a timeline — it follows the band.

Live AudioAnalog Input
// Cohort Evolution

Patagonia → Rock Stage

Daydream Cohort 2: NOMAD Dreamscapes. A solar-powered motorhome traversing Patagonia. The input was nature — lakes, mountains, glaciers. The system processed landscapes into generative textures from inside a moving vehicle.

Cohort 3: Perro Andaluz. Same TouchDesigner → Scope → Resolume pipeline. But now the input is a kinetic live rock band. The slider configurations transferred — but the intensity multiplied by ten. Proof that the architecture is input-agnostic.

C2: NatureC3: Rock
// Paulina Olivero

Artistic Direction

Co-director of "Relatos Salvajes" and "Mi vecino de al lado." Paulina parameterized cinematic vision into the UI. Every track acts as a scene — guided by carefully handcrafted prompt sequencing "por tiradas."

"Treat every kick drum like a jump cut. Treat every song like a scene." Her filmic approach structures the chaos of real-time generation into intentional narrative arcs.

CinemaDirection
// Heavy Distortion

Aggressive Pulse — Track Profile

Red-dominant palette. Maximum prompt rate — every kick drum triggers a full frame mutation. The LoRA was trained on textures of underground Argentine rock: concrete walls, peeling paint, rusted metal. Identidad de barro.

Denoise strength pushed to maximum. Structure weight low. The AI is unleashed — barely constrained by Canny edges. The result is controlled chaos that mirrors the sound.

MAX
Prompt rate
0.85
Denoise
LOW
Structure
// Atmospheric

Sustained Hallucination — Track Profile

Blue-to-purple gradient shift. Slower denoise decay — the visual transformation lags intentionally behind the audio, creating a dream-like persistence. Guitar sustain stretches the hallucination across frames.

LoRA trained on deep ocean, aurora borealis, bioluminescence. ControlNet Depth is dominant — preserving spatial depth while allowing color to bleed freely.

SLOW
Decay rate
0.45
Denoise
HIGH
Depth weight
// Cinematic

Scene Tension — Track Profile

Deep blacks with cyan accent cuts. Paulina's direction — every silence is a long take, every explosion is a jump cut. ControlNet Canny locked tight on the singer's silhouette.

Prompt sequences switch between "stillness" and "eruption" descriptors. The transition speed maps directly to the song's dynamic range. Cinema grammar applied to real-time inference.

LOCKED
Canny
VAR
Denoise
HIGH
Contrast
// Natural Fluids

Winamp Memory — Track Profile

Pure TouchDesigner fluid simulations. No AI generation for this track — the visual is born entirely from audio-reactive 2D fluid dynamics. Turbulence and velocity driven by specific frequency bands.

It harks back to early Winamp visualizations — organic randomness you watched as a kid, but controlled with 32-bit floating-point precision on cloud infrastructure. The purest audio-to-image truth.

NONE
AI gen
PURE
TD fluid
32bit
Precision