GitHub - interrupt21h/tubeviz: AI-directed, beat-synchronized music video generation from YouTube footage, with semantic analysis, interactive editing, and GPU-accelerated rendering.

AI-directed, beat-synchronized music video generation from a curated video library.

tubeviz turns a music track and a library of source footage into a beat-aligned, AI-directed music video. It can acquire and curate footage, detect scenes, analyze music, plan an edit, apply visual treatments, preview the result interactively, and render a finished video.

Tubeviz provides both a browser-based Studio workflow and a full command-line interface. Both operate on the same persistent media library and the same directed timeline format.

Sample videos

Complete videos produced with tubeviz:

• Tubeviz - Moby - God Moving Over the Face of the Waters
• Tubeviz - Empire of the Sun - Walking on a Dream
• Tubeviz - Night Tapes - Drifting
• Tubeviz — Andrew Bayer feat. Alison May — Open End Resource (OCULA Remix)
• Tubeviz — Step It Up — Stereo MC's

Architecture

Tubeviz is built around a persistent clip library and a versioned directed timeline. The timeline is a JSON production plan that separates analysis and editorial decisions from playback and rendering. Once a timeline has been created, it can be previewed, replanned, materialized, or rendered without repeating the complete ingest and analysis workflow.

%%{init: {
  "theme": "base",
  "themeVariables": {
    "background": "#0b1020",
    "primaryColor": "#151b2f",
    "primaryTextColor": "#e8edf7",
    "primaryBorderColor": "#5ee7df",
    "lineColor": "#71809f",
    "secondaryColor": "#1a2036",
    "tertiaryColor": "#11172a",
    "clusterBkg": "#0d1325",
    "clusterBorder": "#34405f",
    "fontFamily": "Inter, ui-sans-serif, system-ui, sans-serif",
    "fontSize": "15px"
  },
  "flowchart": {
    "curve": "basis",
    "htmlLabels": false,
    "nodeSpacing": 48,
    "rankSpacing": 58,
    "diagramPadding": 20
  }
}}%%

flowchart TB

    subgraph INPUTS["INPUTS"]
        BRIEF["✦ Visual brief<br/>or search terms"]
        URLS["▶ Curated<br/>YouTube URLs"]
        AUDIO["♫ Music track"]
    end

    subgraph TV["TUBEVIZ"]
        STUDIO["Studio + CLI"]

        INGEST["Acquisition<br/>& ingest"]
        LIB[("Persistent<br/>clip library")]

        PLAN["Music analysis<br/>& visual direction"]
        TL[("Directed<br/>timeline JSON")]

        PREVIEW["Interactive<br/>preview"]
        RENDER["Browser or<br/>native renderer"]

        AUDIOPATH["♫ Audio source"]
    end

    OUTPUT["◆ Encoded music video"]

    BRIEF --> STUDIO
    URLS --> STUDIO
    AUDIO --> STUDIO

    STUDIO --> INGEST
    INGEST --> LIB

    STUDIO --> PLAN
    LIB --> PLAN

    PLAN --> TL

    TL --> PREVIEW
    TL --> RENDER

    LIB -.-> PREVIEW
    LIB -.-> RENDER

    AUDIO --> AUDIOPATH
    AUDIOPATH -.-> PLAN
    AUDIOPATH -.-> PREVIEW
    AUDIOPATH -.-> RENDER

    RENDER --> OUTPUT

    classDef input fill:#131a2d,stroke:#6878a5,stroke-width:1.5px,color:#e7ebf5;
    classDef process fill:#172036,stroke:#5ee7df,stroke-width:1.8px,color:#f4f8ff;
    classDef intelligence fill:#221b3d,stroke:#b794f6,stroke-width:2px,color:#f5efff;
    classDef storage fill:#101d29,stroke:#38bdf8,stroke-width:2px,color:#eaf9ff;
    classDef render fill:#252035,stroke:#f0abfc,stroke-width:2px,color:#fff1ff;
    classDef output fill:#103026,stroke:#6ee7b7,stroke-width:2.5px,color:#eafff5;
    classDef auxiliary fill:#111827,stroke:#64748b,stroke-width:1.3px,color:#cbd5e1;

    class BRIEF,URLS,AUDIO input;
    class STUDIO,INGEST,PREVIEW process;
    class PLAN intelligence;
    class LIB,TL storage;
    class RENDER render;
    class OUTPUT output;
    class AUDIOPATH auxiliary;

    style INPUTS fill:#0d1325,stroke:#34405f,stroke-width:1px,color:#94a3b8;
    style TV fill:#0d1325,stroke:#435170,stroke-width:1.5px,color:#cbd5e1;

    linkStyle default stroke:#71809f,stroke-width:1.5px;

Tubeviz combines several kinds of information when constructing an edit:

• beat, bar, onset, tempo, section, and phrase structure;
• optional learned audio representations and semantic audio analysis;
• detected source scenes, motion, complexity, palette, and temporal visual features;
• optional OpenCLIP embeddings and AI-generated visual descriptions;
• transition quality, novelty, clip reuse, curation preferences, and sequence continuity;
• transforms, compositing, temporal effects, vector treatments, and optional codec-space effects.

Source video remains the primary visual material. Effects are scheduled around the footage rather than replacing it with a standalone audio-reactive shader.

Contents

• Quick start
• Requirements
• Installation
• Studio
• Settings and AI configuration
• Building a footage library
• Curating the library
• Analyzing music and creating a timeline
• Visual direction and effects
• Previewing
• Rendering
• Hardware acceleration
• FFglitch codec-space effects
• Library layout
• Recommended workflow
• Troubleshooting
• Command overview
• License

Quick start

For most users, Studio is the easiest way to work with tubeviz.

python -m venv .venv
source .venv/bin/activate
pip install -e '.[semantic,audio-ai,render]'

tubeviz gui \
  --project-root "$PWD" \
  --library ./library

Studio opens in the browser, normally at:

A typical project flow is:

1. Build or import a footage library
2. Review, trim, reject, and enhance clips
3. Analyze a music track and create a directed timeline
4. Preview and adjust the edit
5. Render the final video

The same workflow can be performed entirely from the command line:

# Acquire footage.
tubeviz ingest \
  --terms search_terms.txt \
  --library ./library \
  --results-per-term 10

# Analyze the track and construct an edit.
tubeviz analyze audio/song.mp3 \
  --library ./library \
  --semantic \
  --output timelines/song.json

# Preview it.
tubeviz serve timelines/song.json \
  --library ./library \
  --audio audio/song.mp3

# Render it.
tubeviz render timelines/song.json \
  --library ./library \
  --audio audio/song.mp3 \
  --output output/song.mp4 \
  --backend auto

Requirements

Core requirements

• Python 3.11+
• FFmpeg and ffprobe
• yt-dlp, installed with the Python package

Optional components

Linux system packages

Arch Linux / CachyOS

sudo pacman -S --needed \
  base-devel cmake pkgconf ffmpeg libplacebo vulkan-icd-loader \
  vulkan-tools curl unzip chromium

Debian / Ubuntu

sudo apt update
sudo apt install -y \
  build-essential cmake pkg-config ffmpeg curl unzip \
  libavformat-dev libavcodec-dev libavutil-dev libswscale-dev \
  libplacebo-dev libvulkan-dev vulkan-tools

Install Chrome/Chromium separately, or use Playwright's managed Chromium build with playwright install chromium. Distribution packaging for Chromium differs between Debian and Ubuntu.

The compiler, CMake, FFmpeg development headers, libplacebo, and Vulkan packages are needed only for the native renderer and its optional GPU path. Browser rendering does not depend on the native renderer.

Installation

Clone the repository and install the base package:

git clone https://github.com/interrupt21h/tubeviz.git tubeviz
cd tubeviz
python -m venv .venv
source .venv/bin/activate
pip install -e .

Install all commonly used optional features:

pip install -e '.[semantic,audio-ai,render]'

The extras can also be installed independently:

If you want Playwright to manage its own Chromium build:

playwright install chromium

Verify the installation:

tubeviz --help
ffmpeg -version

For optional acceleration features, also run:

tubeviz native doctor
tubeviz audio-ai doctor
tubeviz music-ai doctor
tubeviz codec doctor

Studio

Studio is the easiest way to work with tubeviz interactively. It organizes the workflow into focused workspaces while keeping the current project paths and library state available as you move between stages.

tubeviz gui \
  --project-root /path/to/tubeviz-project \
  --library /path/to/tubeviz-project/library

Other launch examples:

tubeviz gui --library ./library --port 8095
tubeviz gui --library ./library --no-open
tubeviz gui --host 0.0.0.0 --port 8090

When binding Studio to a non-loopback address, place it behind appropriate network controls if other systems can reach it.

Project

The Project workspace defines the files and directories used by the current production: the footage library, music track, directed timeline, and rendered output. It also provides a compact overview of project state and shortcuts into the rest of the workflow.

Ingest

The Ingest workspace is used to build and expand the footage library. Choose the acquisition method that matches the material you have:

• AI Discovery starts from a visual brief and searches for footage that fits it;
• Search Terms uses explicit visual concepts or a search-term file;
• YouTube URLs imports known sources directly.

Download quality, source filtering, scene detection, media preparation, and optional analysis controls are shared across the ingest workflow.

Library

The Library workspace is where source footage is reviewed and curated before it is used in an edit. It supports:

• filtering and browsing clips;
• video playback;
• non-destructive In/Out trimming;
• rejection and restoration;
• permanent deletion;
• tags and output-pool selection;
• scene, visual, semantic, and AI metadata inspection.

The clip detail view keeps playback and trim controls together so usable source ranges can be adjusted while watching the footage.

Timeline

The Timeline workspace combines music analysis, editorial direction, interactive preview, and inspection of the generated edit. The timeline beneath the preview shows the structure of the production over time, including musical sections, energy, beats, selected clips, transitions, creative effects, vector or codec treatments, and AI Director events when used.

Clicking timeline items exposes details about the selected shot or event, and the preview can be scrubbed alongside the timeline or opened in a separate window.

Render

The Render workspace turns a directed timeline into the finished video. It provides the output resolution, frame rate, codec, quality, backend, and acceleration controls needed for browser or native rendering.

For normal use, the automatic backend and acceleration settings are a good starting point; the explicit controls are available when a particular renderer or hardware path is required.

Jobs

Long-running work is tracked in the Jobs workspace. It shows current and previous jobs, progress, elapsed time, estimated completion when available, complete arguments, captured output, and cancellation controls for supported operations.

A compact activity bar also reports foreground work while you use other Studio tabs.

Settings

The Settings workspace contains application-wide AI and model configuration, including the master AI switch, OpenAI-compatible endpoint and model settings, credentials, image analysis limits, timeouts, and video-understanding options.

Advanced

The Advanced workspace provides the Command Center for users who want direct access to tubeviz commands and their full argument sets without leaving Studio. It is useful for specialized workflows and for reproducing a Studio operation from the command line.

For exact command syntax, the CLI help is the authoritative reference:

tubeviz --help
tubeviz analyze --help
tubeviz render --help

Settings and AI configuration

Studio's Settings workspace is the central configuration surface for optional learned and API-backed capabilities.

Configuration includes:

• a master AI-feature switch;
• OpenAI API key;
• OpenAI-compatible base URL;
• shared OpenAI model;
• Hugging Face token;
• image detail and frame budget for visual description;
• API timeouts;
• storyboard/video-understanding controls.

Settings are stored outside the repository at:

~/.config/tubeviz/config.json

or, when XDG_CONFIG_HOME is set:

$XDG_CONFIG_HOME/tubeviz/config.json

Set TUBEVIZ_CONFIG to use a different configuration file.

Environment variables can also provide credentials:

export OPENAI_API_KEY='...'
export HF_TOKEN='hf_...'

Hugging Face authentication

Public models can often be downloaded without a token, but a Hugging Face token is useful for gated models, authenticated access, and rate limits.

A read token is sufficient for model downloads.

AI video understanding

When enabled, tubeviz can analyze sampled frames from clips and store descriptions that cover subjects, actions, location, camera language, lighting, palette, texture, mood, editing utility, and scene-level characteristics.

This metadata supplements local measurements and OpenCLIP embeddings; it does not replace deterministic scene analysis.

Existing clips can be analyzed from the CLI:

tubeviz library ai-describe --library ./library

Limit the operation or target a specific clip:

tubeviz library ai-describe --library ./library --limit 10
tubeviz library ai-describe --library ./library --clip-id 42 --force

Building a footage library

The footage library is persistent. Downloaded media, scene indexes, thumbnails, embeddings, trim ranges, curation state, provenance, and derived analysis can be reused across many songs and timelines.

Theme-first acquisition

A visual brief lets tubeviz generate a diverse set of searches and evaluate candidates before committing to full downloads.

tubeviz ingest \
  --visual-brief 'A nocturnal electronic dream: fluorescent city motion, abstract machinery, wet streets, refracted glass, underground dance energy, cinematic movement. Avoid title cards, logos, talking heads, tutorials, and static footage.' \
  --audio audio/song.mp3 \
  --library ./library \
  --target-clips 40 \
  --acquisition-query-count 24 \
  --preview-gate \
  --preview-samples 4 \
  --preview-seconds 4 \
  --min-video-fitness 0.18 \
  --auto-trim

The preview gate evaluates short samples before a full download. It can reject footage with insufficient motion or useful visual activity, excessive text overlays, dominant talking heads, or other characteristics that make it a poor fit for music-video editing.

Search-term acquisition

Create a text file containing one visual concept per line:

underground techno warehouse strobe
laser tunnel rave crowd
analog CRT glitch surveillance
industrial machinery sparks
cyberpunk city rain neon
abstract liquid chrome macro
satellite earth night timelapse
high speed train tunnel POV

Then ingest it:

tubeviz ingest \
  --terms search_terms.txt \
  --library ./library \
  --results-per-term 10 \
  --cookies-from-browser chrome

Semantic discovery and scene embedding can be enabled when the semantic extra is installed:

tubeviz ingest \
  --terms search_terms.txt \
  --library ./library \
  --results-per-term 10 \
  --ai-discovery \
  --ai-query-expansion \
  --ai-query-count 8 \
  --ai-candidates-per-term 100 \
  --ai-device auto \
  --ai-index-scenes

Useful ingest controls include:

Use tubeviz ingest --help for the complete acquisition and filtering controls.

Manual URL ingest

Known YouTube sources can be added directly:

tubeviz ingest-url \
  'https://www.youtube.com/watch?v=VIDEO_ID' \
  --library ./library

Multiple URLs are accepted:

tubeviz ingest-url URL1 URL2 URL3 \
  --library ./library \
  --term hand-picked

Manual ingest uses the same library pipeline as searched footage, including metadata, duplicate checks, scene indexing, thumbnails, visual analysis, and optional semantic indexing.

Curating the library

Inspect basic library state:

tubeviz library stats --library ./library
tubeviz library list --library ./library --limit 50
tubeviz library list --library ./library --status ready
tubeviz library show VIDEO_ID --library ./library

Reject and restore

Rejecting a clip is non-destructive and prevents it from being selected for future output:

tubeviz library reject VIDEO_ID \
  --library ./library \
  --reason 'static talking-head footage'

Restore it later:

tubeviz library restore VIDEO_ID --library ./library

Delete

Permanent deletion removes the clip and its tracked derived assets:

tubeviz library delete VIDEO_ID --library ./library

Use rejection when you may want to reconsider a clip. Use deletion when the source and its generated assets should be removed completely.

Trim usable source ranges

Studio supports non-destructive In/Out marks for clips containing intros, title cards, credits, black leader, or other unusable edge material. The original source remains unchanged; the saved usable range constrains future scene planning.

Build or refresh visual indexes

Local temporal visual features:

tubeviz library visual-index --library ./library

OpenCLIP scene embeddings:

tubeviz library embed --library ./library --device auto

Analyzing music and creating a timeline

The analyze command extracts musical structure and can immediately attach a scene plan from the current library.

A useful general-purpose starting point is:

tubeviz analyze audio/song.mp3 \
  --library ./library \
  --output timelines/song.json \
  --semantic \
  --semantic-device auto \
  --section-bars 8 \
  --max-video-layers 3 \
  --composition-intensity 1.2 \
  --transform-intensity 1.2 \
  --novelty-weight 0.65 \
  --visual-match-weight 1.25 \
  --transition-weight 0.7 \
  --vector-intensity 1.0 \
  --selection-variation 0.30 \
  --min-shot-seconds 0.65 \
  --max-shot-seconds 6.0 \
  --source-excerpt-max-seconds 5.0

The resulting timeline contains musical analysis, visual intent, selected footage, source ranges, transforms, compositing, effect schedules, and provenance required by the preview and rendering paths.

Variable tempo and phrase choreography

Tubeviz can model local tempo instead of assuming one BPM for the entire track. Phrase and section analysis can influence shot duration, visual intensity, anticipation before peaks, and release after them.

Useful controls include:

--section-bars
--tempo-window-seconds
--tempo-smoothing-seconds
--tempo-change-bpm
--trajectory-strength
--anticipation-seconds
--visual-arc-strength

CLAP audio semantics

CLAP can classify sliding windows of the music and influence scene selection and visual treatment:

tubeviz audio-ai doctor --device auto

Enable it during analysis:

tubeviz analyze audio/song.mp3 \
  --library ./library \
  --output timelines/song.json \
  --semantic \
  --audio-ai \
  --audio-ai-device auto \
  --audio-ai-window 8 \
  --audio-ai-hop 4 \
  --audio-visual-match-weight 1.1

MERT music representations

MERT can add learned structural representations for novelty and musical change:

tubeviz music-ai doctor --device auto

Enable it with:

tubeviz analyze audio/song.mp3 \
  --library ./library \
  --music-ai \
  --music-ai-device auto \
  --output timelines/song.json

AI director

When configured, the optional AI director can provide whole-song visual direction and bounded editorial guidance while deterministic validation continues to control actual scene IDs, durations, source availability, and renderer capabilities.

tubeviz analyze audio/song.mp3 \
  --library ./library \
  --semantic \
  --ai-director \
  --output timelines/song.json

Studio can supply the configured API endpoint, model, and key automatically.

Alternate cuts

Use a deterministic seed for reproducible variation:

tubeviz analyze audio/song.mp3 \
  --library ./library \
  --selection-seed 12345 \
  --selection-variation 0.35 \
  --output timelines/song-alt.json

Or request a fresh seed:

tubeviz analyze audio/song.mp3 \
  --library ./library \
  --reshuffle \
  --output timelines/song-alt.json

Visual direction and effects

Tubeviz schedules effects according to musical and visual context. The renderer can mix ordinary source transforms with semantic/temporal effects, vector treatments, layered composition, and optional codec-space effects.

Video-first transforms

Typical source treatments include:

• crop, scale, pan, rotation, and virtual-camera motion;
• flow and harmonic warp;
• pseudo-depth parallax;
• RGB displacement;
• bloom, streaks, scanlines, noise, and vignette;
• palette and contrast direction;
• beat-local deformation with musically selected push, pinch, shear, twist, wave, saddle, lens, and spiral motion;
• multi-source composition and crossfades.

The goal is to preserve source identity while creating motion and continuity that follow the song.

Semantic temporal effects

The creative renderer can combine persistent temporal state with source-derived visual features. Depending on the planned shot, treatments can include:

• feedback and frame echo;
• motion trails;
• slit-scan/history effects;
• pixelation and posterization;
• solarization and edge extraction;
• glitch and block displacement;
• VHS-style tracking;
• ripple and tempo deformation;
• sparse depth, flow, and focal-point treatments.

Effect strength is controlled by the timeline and the analysis settings rather than by a single global preset.

Vector effects

Vector treatments derive structure from source imagery and motion. They can be used for sparse outlines, motion-linked geometry, motif memory, and scene-responsive graphical accents.

Control their overall contribution with:

--vector-effects / --no-vector-effects
--vector-intensity

Composition

Tubeviz can plan multiple simultaneous source layers when appropriate:

--max-video-layers
--composition-intensity
--clip-opacity
--scene-crossfade

Higher values can produce denser edits, but one or two strong source layers are often more readable than continuous maximum-density composition.

Previewing

Serve an existing timeline:

tubeviz serve timelines/song.json \
  --library ./library \
  --audio audio/song.mp3

Studio also provides a managed preview directly in the Timeline workspace.

The interactive preview uses a responsive pipeline by default. Tubeviz prepares a lightweight 720p/30fps preview-media cache independently of final-render media, adapts its internal render resolution and layer count to the measured frame budget, and caps live presentation near 30 fps. When WebGPU is available, source video surfaces are composited directly as GPU external textures before the fused post-processing pass, avoiding the full-frame Canvas2D composition copy. Canvas2D remains the compatibility path.

Studio's Preview decode control normally chooses HTML video for the direct WebGPU path and can use the worker WebCodecs decoder when Canvas fallback benefits from moving decode work off the main thread. Full fidelity preview remains available when exact browser-effect inspection matters more than interaction speed. Native/final rendering never uses these preview-quality reductions.

For best browser GPU support:

• use a current Chrome or Chromium build;
• use a secure context or loopback address;
• keep the browser's hardware acceleration enabled;
• confirm the browser exposes WebGPU;
• avoid remote-display configurations that disable GPU access.

The preview HUD reports the active rendering path so it is possible to distinguish WebGPU rendering from Canvas2D fallback.

Rendering

Tubeviz offers two offline rendering backends:

• native — C++/FFmpeg renderer with optional libplacebo/Vulkan effects and CUDA/NVDEC source decode;
• browser — deterministic browser renderer using WebCodecs/WebGPU when available, with compatibility fallbacks.

auto prefers the native backend when a usable native renderer is available.

Automatic backend

tubeviz render timelines/song.json \
  --library ./library \
  --audio audio/song.mp3 \
  --output output/song.mp4 \
  --backend auto \
  --width 1920 \
  --height 1080 \
  --fps 30 \
  --crf 20

Native backend

Build it:

tubeviz native build --clean

Inspect the local toolchain and native build:

Render explicitly with the native backend:

tubeviz render timelines/song.json \
  --library ./library \
  --audio audio/song.mp3 \
  --output output/song.mp4 \
  --backend native \
  --width 1920 \
  --height 1080 \
  --fps 30 \
  --video-codec libx264 \
  --crf 20 \
  --native-preset veryfast \
  --native-decoder-cache 16 \
  --native-threads 0 \
  --native-gpu auto \
  --native-hwdecode auto \
  --native-build-if-missing

Native GPU controls:

Browser backend

tubeviz render timelines/song.json \
  --library ./library \
  --audio audio/song.mp3 \
  --output output/song-browser.mp4 \
  --backend browser \
  --width 1920 \
  --height 1080 \
  --fps 30 \
  --browser-transport auto \
  --browser-gpu auto \
  --browser-source-decode auto

Browser acceleration controls:

Use explicit webgpu or webcodecs modes when diagnosing acceleration. Use auto for normal operation.

Hardware acceleration

Tubeviz can use several independent acceleration paths. Availability of one does not imply availability of the others.

%%{init: {
  "theme": "base",
  "themeVariables": {
    "background": "#ffffff",
    "primaryTextColor": "#f8fafc",
    "textColor": "#f8fafc",
    "lineColor": "#64748b",
    "edgeLabelBackground": "#1e293b",
    "fontFamily": "Inter, ui-sans-serif, system-ui, sans-serif",
    "fontSize": "16px"
  },
  "themeCSS": "
    .edgeLabel {
      background-color: #1e293b !important;
      color: #f8fafc !important;
      border-radius: 4px;
      padding: 2px 6px;
    }
    .edgeLabel p {
      color: #f8fafc !important;
      background-color: #1e293b !important;
    }
    .label text,
    .nodeLabel {
      font-weight: 600;
    }
  ",
  "flowchart": {
    "curve": "basis",
    "nodeSpacing": 45,
    "rankSpacing": 60
  }
}}%%

flowchart LR
    SRC["Source video"] --> DECODE{"Source decode"}

    DECODE -->|"CUDA / NVDEC"| CUDA["NVIDIA decode"]
    DECODE -->|"Software"| CPUDEC["CPU decode"]

    CUDA --> FX{"Creative FX"}
    CPUDEC --> FX

    FX -->|"libplacebo / Vulkan"| VK["GPU effects"]
    FX -->|"CPU"| CPUFX["CPU effects"]

    VK --> ENC["FFmpeg encoder"]
    CPUFX --> ENC

    classDef source fill:#172033,stroke:#64748b,stroke-width:2px,color:#f8fafc
    classDef decision fill:#4c1d5f,stroke:#a855f7,stroke-width:2px,color:#ffffff
    classDef gpu fill:#06384a,stroke:#06b6d4,stroke-width:2.5px,color:#ecfeff
    classDef cpu fill:#1e293b,stroke:#64748b,stroke-width:2px,color:#f8fafc
    classDef output fill:#641747,stroke:#ec4899,stroke-width:2.5px,color:#ffffff

    class SRC source
    class DECODE,FX decision
    class CUDA,VK gpu
    class CPUDEC,CPUFX cpu
    class ENC output

    linkStyle 0 stroke:#64748b,stroke-width:2px
    linkStyle 1,3,5,7 stroke:#06b6d4,stroke-width:3px
    linkStyle 2,4,6,8 stroke:#64748b,stroke-width:2px

Native diagnostics

Also test the underlying runtimes directly when diagnosing GPU availability:

vulkaninfo --summary
nvidia-smi
ffmpeg -hide_banner -hwaccels

libplacebo being installed means the native renderer can be built with libplacebo support; it does not guarantee that Vulkan can create a usable GPU device at runtime. Likewise, cuda appearing in ffmpeg -hwaccels means the FFmpeg build advertises CUDA support; actual source decoding still depends on the driver, codec, and runtime device.

WSL notes

Under WSL, CUDA/NVDEC and Vulkan are separate paths:

• CUDA/NVDEC uses the NVIDIA Windows driver exposed to WSL;
• Vulkan requires a Vulkan implementation visible inside the Linux environment;
• libplacebo can only use Vulkan devices exposed by that Linux Vulkan stack.

If vulkaninfo --summary reports only a CPU renderer such as llvmpipe, native libplacebo effects will not use the GPU. Resolve Vulkan device visibility before troubleshooting tubeviz itself.

FFglitch codec-space effects

FFglitch is optional. Tubeviz uses the external ffedit program to materialize true codec-space motion-vector effects. Other FFglitch tools are not required for normal tubeviz operation.

Check availability:

Linux x86-64 installation

Download the FFglitch release archive from the official FFglitch site, extract ffedit, and place it on PATH. A user-local installation can use ~/.local/bin:

mkdir -p ~/.local/bin
# Copy the extracted ffedit binary into ~/.local/bin/ffedit
chmod 0755 ~/.local/bin/ffedit
export PATH="$HOME/.local/bin:$PATH"

Verify it:

command -v ffedit
ffedit -h | head -40
tubeviz codec doctor

Schedule codec effects

Codec-space effects can be included while analyzing a track:

tubeviz analyze audio/song.mp3 \
  --library ./library \
  --codec-glitch musical \
  --codec-glitch-intensity 0.65 \
  --output timelines/song.json

Available scheduling profiles are:

off
subtle
musical
aggressive

Materialize codec shots

tubeviz codec materialize timelines/song.json \
  --library ./library

The rendered or previewed timeline can then reference the materialized codec-shot cache.

If FFglitch is unavailable, ordinary analysis, preview, vector effects, and rendering remain usable.

Library layout

A tubeviz library is self-contained and reusable. The exact set of derived directories can vary with enabled features, but the high-level structure is:

library/
├── metadata.sqlite3
├── originals/
├── normalized/
├── thumbnails/
├── metadata/
├── transforms/                 # created when transform materialization is used
├── codec-glitch/               # created when codec effects are materialized
└── browser-webcodecs-cache/    # created by accelerated browser rendering

SQLite stores clip metadata, discovery provenance, scenes, duplicate relationships, trim ranges, tags, curation state, visual features, AI analysis, and scene embeddings. Generated directories are created only when the corresponding workflow needs them.

Tags and output pool

User tags are separate from acquisition/search terms. Search terms record how footage entered the library; tags are for organizing footage for reuse. Studio can also maintain an output pool that restricts planning to selected ready clips.

With an empty output pool, all eligible ready clips can be selected. When clips are marked for output, analysis and replanning use only those marked clips. Clearing the pool returns planning to the full ready library.

Do not treat generated cache directories as irreplaceable source media. The SQLite library database and original source assets are the important persistent state.

Recommended workflow

The following workflow works well for a full-length electronic track.

1. Build a diverse library

Start with a strong visual brief or a varied set of search concepts. Favor footage with movement, changing composition, useful camera motion, and minimal persistent text.

tubeviz ingest \
  --visual-brief 'Kinetic nocturnal electronic imagery with industrial motion, neon infrastructure, crowds, transit, abstract macro textures, and cinematic movement.' \
  --audio audio/song.mp3 \
  --library ./library \
  --target-clips 50 \
  --preview-gate \
  --auto-trim

2. Curate it

Use Studio's Library view to reject weak material and trim intros/outros. A smaller, high-quality pool usually produces better edits than a large pool full of visually static sources.

3. Add semantic indexes

tubeviz library visual-index --library ./library
tubeviz library embed --library ./library --device auto

Optionally add AI visual descriptions from Studio or:

tubeviz library ai-describe --library ./library

4. Analyze the song

tubeviz analyze audio/song.mp3 \
  --library ./library \
  --output timelines/song.json \
  --semantic \
  --semantic-device auto \
  --audio-ai \
  --audio-ai-device auto \
  --section-bars 8 \
  --max-video-layers 3 \
  --composition-intensity 1.2 \
  --transform-intensity 1.2 \
  --novelty-weight 0.65 \
  --visual-match-weight 1.25 \
  --transition-weight 0.7 \
  --vector-intensity 1.0 \
  --selection-variation 0.30 \
  --min-shot-seconds 0.65 \
  --max-shot-seconds 6 \
  --source-excerpt-max-seconds 5

Treat these values as a starting point rather than a required preset. Different music benefits from different shot lengths, layer density, novelty, and effect intensity.

5. Preview and iterate

tubeviz serve timelines/song.json \
  --library ./library \
  --audio audio/song.mp3

If the material is good but the cut is not, create an alternate selection before changing the visual-effect stack:

tubeviz analyze audio/song.mp3 \
  --library ./library \
  --semantic \
  --reshuffle \
  --output timelines/song-alt.json

6. Render

tubeviz render timelines/song.json \
  --library ./library \
  --audio audio/song.mp3 \
  --output output/song.mp4 \
  --backend auto \
  --width 1920 \
  --height 1080 \
  --fps 30 \
  --crf 20

Troubleshooting

YouTube returns 403 / Forbidden

Use current yt-dlp and, when necessary, browser cookies:

pip install -U yt-dlp

tubeviz ingest \
  --terms search_terms.txt \
  --library ./library \
  --cookies-from-browser chrome

If the failure is specific to one video, test it directly with yt-dlp before changing tubeviz settings.

Ingest appears slow

Large source files, high-resolution downloads, scene detection, preview sampling, embedding generation, and AI description can each add significant work.

Useful checks:

tubeviz library stats --library ./library

Use Studio's Jobs panel to identify the active stage. To isolate bottlenecks, temporarily disable optional analysis or reduce the candidate pool rather than assuming the download itself is stalled.

Native renderer cache or version mismatch

The native executable is built into a cache directory. If the Python CLI and native binary appear out of sync, rebuild it cleanly:

tubeviz native build --clean
tubeviz native doctor

You can also inspect the executable reported by native doctor and run it directly with --version.

Native Vulkan effects are unavailable

Start with:

tubeviz native doctor
vulkaninfo --summary

Check both of the following:

1. libplacebo is visible to the build through pkg-config;
2. Vulkan can create a suitable runtime device.

For example:

pkg-config --modversion libplacebo
vulkaninfo --summary

If Vulkan sees only a CPU renderer, native GPU Creative FX will not be available even though libplacebo is installed.

Use this render mode to make Vulkan failure explicit while diagnosing it:

Return to:

for normal operation with CPU fallback.

CUDA/NVDEC is not being used

A native log line such as:

decoder_open=... hw=software

means the source is being decoded on the CPU.

Test CUDA independently with the same source file:

nvidia-smi

ffmpeg -hide_banner -loglevel verbose \
  -hwaccel cuda \
  -hwaccel_output_format cuda \
  -i /path/to/source.mp4 \
  -frames:v 30 \
  -f null -

If that command fails, fix the FFmpeg/driver/runtime problem before troubleshooting the tubeviz decoder.

Use this render mode to require CUDA during diagnosis:

Browser preview is using Canvas2D instead of WebGPU

Check:

• Chrome/Chromium is current;
• browser hardware acceleration is enabled;
• the preview is served from loopback or another secure context;
• navigator.gpu is available in DevTools;
• the preview HUD reports the WebGPU initialization reason;
• the GPU/driver is not blocked by the browser.

For normal interactive use, automatic fallback keeps the preview usable. For browser offline-render diagnostics, require WebGPU explicitly:

Browser rendering is slow

Use automatic accelerated paths first:

--browser-transport auto
--browser-gpu auto
--browser-source-decode auto

If the browser cannot use WebCodecs or WebGPU, the fallback paths can be substantially slower at 1080p. Compare with the native backend when available.

Studio Play reports no media

Confirm the clip has usable local media and inspect it from the CLI:

tubeviz library show VIDEO_ID --library ./library

If the clip was only partially ingested, reprocess or delete/reingest it as appropriate.

Confirming which installation is running

command -v tubeviz
python -c 'import tubeviz, inspect; print(inspect.getfile(tubeviz))'
tubeviz --help

When running from a local checkout, activate the intended virtual environment before launching Studio or rendering.

Command overview

Top-level commands:

Use the built-in help for the complete current option set:

tubeviz --help
tubeviz ingest --help
tubeviz ingest-url --help
tubeviz library --help
tubeviz analyze --help
tubeviz serve --help
tubeviz render --help
tubeviz codec --help
tubeviz native --help

License

Tubeviz is licensed under the Apache License 2.0. See LICENSE and NOTICE.

Third-party software, models, and media

Tubeviz integrates with external software and optional model providers, each of which is subject to its own license and terms. Examples include FFmpeg, yt-dlp, Playwright, OpenCLIP, PyTorch, Transformers, libplacebo, Vulkan implementations, and FFglitch.

Users are responsible for ensuring that source media is acquired and used in accordance with applicable licenses, copyright law, platform terms, and any other relevant permissions.

Author

Tubeviz was created and is maintained by Scott Muller.

GitHub: @interrupt21h