media-processing

Use input-side seeking with stream copy: ffmpeg -ss 00:01:30 -i in.mp4 -to 00:02:00 -c copy -avoid_negative_ts make_zero cut.mp4. The cut lands on the nearest keyframe, so report the actual cut point or re-encode when frame accuracy matters.

scale=1280:-1 preserves aspect ratio exactly and can compute an odd height, which libx264 with yuv420p rejects. Use scale=1280:-2 instead, which rounds to an even dimension and removes the entire failure class.

Use an image tool, never ffmpeg, for picture-in picture-out work. ffmpeg discards EXIF blocks and its autorotate behaviour varies by build, while vips, sharp, ImageMagick, and Pillow apply orientation and ICC conversion as explicit documented calls.

Run loudnorm in two passes: first measure with print_format=json, then apply with the measured values and linear=true. Always follow it with aresample, because loudnorm silently resamples its output to 192 kHz.

Each rendition picks its own scene-change keyframes, leaving no common splice point. Fix it by pinning one GOP across all renditions with -g and -keyint_min at framerate times segment duration plus -sc_threshold 0.

Do not use it for model-based generation, editing, or transcription of media, for framework-owned image pipelines like next/image, or for office documents such as PDF and docx. It only executes deterministic transformations of existing media files.