ascii-video

Sample frames from the source video, map per-cell luminance to characters from a density-sorted palette, and color cells from the original pixel values. Frames are composited with pre-rasterized font bitmaps and piped to ffmpeg for H.264 encoding.

Analyze the audio with FFT to extract frequency bands, RMS energy, spectral centroid, and beat events per frame. Drive value field generators, particle systems, and beat-reactive shaders from those features, then encode frames to MP4 with the audio muxed.

No GPU is required. Rendering runs on CPU with NumPy-vectorized effects and a per-cell compositing loop, using concurrent.futures workers for parallel encoding. Hardware detection adapts resolution and worker count automatically.

ASCII characters are small bright marks on black, so most frame pixels are background and linear brightness multipliers clip highlights. Use percentile-based adaptive tonemapping with gamma around 0.75 instead of multiplying the canvas.

Setting stderr=subprocess.PIPE on a long-running ffmpeg process deadlocks once the buffer fills at about 64KB. Redirect stderr to a log file instead and keep stdin as the raw frame pipe.

The per-cell font bitmap compositing loop runs at roughly 100-150ms per frame and is the bottleneck, so a 5-minute 1080p video takes about 17 minutes on 8 workers. Python loops over rows and columns in effect functions will further degrade performance.