This commit is contained in:
Andras Schmelczer 2026-05-11 21:38:26 +01:00
parent 9248e26af2
commit f2a2651b8a
95 changed files with 3993 additions and 1471 deletions

View file

@ -1,6 +1,11 @@
#!/usr/bin/env bash
#
# End-to-end re-render of the dashboard demo video.
# End-to-end re-render of the dashboard demo videos.
#
# All per-storyboard knobs (aspect, fps, bitrate, prompt text, voice persona,
# poster timestamp, brand strings…) live on the Storyboard objects in
# src/storyboard.ts. To add a vertical cut or change the voice, edit that
# file — this script only handles target/auth/transport concerns.
#
# Two targets:
# local (default) — assumes the docker-compose stack on host.docker.internal,
@ -17,7 +22,6 @@
# ./render.sh --no-audio # skip Qwen3-TTS narration; silent MP4
# FORCE_AUTH=1 ./render.sh # same as --fresh-auth
# APP_URL=http://localhost:3001 ./render.sh # override frontend URL
# TTS_SPEAKER=aiden ./render.sh # override CustomVoice speaker
#
# Cred env vars (read for both targets, but prod has no fallback defaults):
# LOGIN_EMAIL, LOGIN_PASSWORD — the dashboard account to record as
@ -48,7 +52,7 @@ case "$TARGET" in
*) echo "Unknown --target: $TARGET (expected: local, prod)" >&2; exit 2 ;;
esac
# -- config (override via env) -------------------------------------------------
# -- environment (target-specific URLs and credentials) ----------------------
if [ "$TARGET" = "prod" ]; then
# Prod serves frontend, /api/*, and /pb/* off the same domain.
export APP_URL="${APP_URL:-https://perfect-postcode.co.uk}"
@ -81,23 +85,6 @@ AUTH_TTL_HOURS="${AUTH_TTL_HOURS:-24}" # re-auth if cache older than this
# the built bundle, so updating this path is what makes the new clip appear
# on the homepage. Override if the dashboard ever moves.
PUBLISH_DIR="${PUBLISH_DIR:-../frontend/public/video}"
# When in the output timeline to grab the poster frame.
# Right-pane inspection (~16s output) is the clearest paused-state preview:
# Manchester map, filters applied, right pane populated, larger narration
# caption visible.
POSTER_TIME_S="${POSTER_TIME_S:-16}"
# Recorder/encoder knobs read by src/config.ts. config.ts treats these as
# required, so they live here (the only entry point) rather than as defaults
# scattered across TS modules. Override per-run via env.
export ASPECT="${ASPECT:-16x9}"
export CAPTURE_SCALE="${CAPTURE_SCALE:-1}"
export WEBM_BITRATE="${WEBM_BITRATE:-$(awk -v s="$CAPTURE_SCALE" 'BEGIN{print (s+0>1)?"18M":"8M"}')}"
export PROMPT_TEXT="${PROMPT_TEXT:-Flats or terraces <£450k, 35 min to Manchester, low crime}"
export AI_ZOOM_SCALE="${AI_ZOOM_SCALE:-2.4}"
export MAX_DURATION_S="${MAX_DURATION_S:-60}"
export MIN_DURATION_S="${MIN_DURATION_S:-10}"
export OUTPUT_FPS="${OUTPUT_FPS:-50}"
FRESH_AUTH="${FORCE_AUTH:-0}"
DO_ENCODE=1
@ -109,7 +96,7 @@ for arg in "${@:-}"; do
--no-encode) DO_ENCODE=0 ;;
--no-audio) DO_AUDIO=0 ;;
-h|--help)
sed -n '3,30p' "$0"
sed -n '3,32p' "$0"
exit 0 ;;
*) echo "Unknown arg: $arg" >&2; exit 2 ;;
esac
@ -207,22 +194,57 @@ else
say "Reusing existing $AUTH_STATE_FILE"
fi
# -- preflight + synth (Qwen3-TTS) -------------------------------------------
# Synth runs BEFORE recording: one batched generate_custom_voice call across
# all cues so the voice stays consistent. The recorder reads
# output/audio/index.json for measured per-cue durations and sizes each
# cue's wall-clock to fit; --no-audio skips synth and the recorder falls
# back to a worst-case estimate.
# -- preflight ---------------------------------------------------------------
# preflight emits per-storyboard narration scripts AND output/storyboards.json
# (the index this script loops over below). Run it BEFORE wiping per-storyboard
# files so we know what slugs to target.
mkdir -p output
# Wipe last run's leaking artifacts so the rename step picks up *this* run.
rm -f output/recording.webm output/recording.mp4 output/page@*.webm output/page@*.webm.untrimmed
rm -f output/narration-script.json output/narration.json
# output/audio/ is preserved; tts/synth.py decides whether the cached WAVs
# still match the script and skips generation when they do.
say "Preflight: emitting narration script"
say "Preflight: emitting narration scripts and storyboard index"
node dist/preflight.js
if [ ! -s output/storyboards.json ]; then
fail "preflight did not produce output/storyboards.json"
fi
# Pull the storyboard slugs out of the index. Use Node so we don't grow a jq
# dependency just for one read.
mapfile -t STORYBOARDS < <(node -e '
const idx = JSON.parse(require("fs").readFileSync("output/storyboards.json","utf8"));
for (const s of idx.storyboards) console.log(s.name);
')
if [ "${#STORYBOARDS[@]}" -eq 0 ]; then
fail "storyboards.json contains no storyboards"
fi
say "Storyboards to render: ${STORYBOARDS[*]}"
# Per-storyboard poster timestamp lookup (slug → seconds), set once so each
# loop body can read it without re-parsing the index.
poster_time_for() {
node -e '
const idx = JSON.parse(require("fs").readFileSync("output/storyboards.json","utf8"));
const sb = idx.storyboards.find(s => s.name === process.argv[1]);
if (!sb) { process.exit(1); }
process.stdout.write(String(sb.posterTimeS));
' "$1"
}
# -- per-storyboard wipe of leaking artefacts --------------------------------
# output/<sb>/audio/ is preserved; tts/synth.py decides whether the cached
# WAVs still match the script and skips generation when they do.
for sb in "${STORYBOARDS[@]}"; do
rm -f "output/$sb/recording.webm" "output/$sb/recording.mp4" \
"output/$sb/page@"*.webm "output/$sb/page@"*.webm.untrimmed \
"output/$sb/recording.raw.webm" "output/$sb/recording.raw.webm.untrimmed" \
"output/$sb/recording.narrated.mp4" "output/$sb/poster.jpg" \
"output/$sb/narration.json"
done
# -- synth (Qwen3-TTS) -------------------------------------------------------
# Synth runs BEFORE recording: one batched generate_voice_clone call per
# storyboard so the voice stays consistent within each video. The recorder
# reads output/<sb>/audio/index.json for measured per-cue durations and
# sizes each cue's wall-clock to fit; --no-audio skips synth and the recorder
# falls back to a worst-case estimate.
if [ "$DO_AUDIO" = "1" ]; then
if ! command -v uv >/dev/null 2>&1; then
fail "uv not on PATH (required for Qwen3-TTS synth). Install uv or rerun with --no-audio."
@ -236,95 +258,103 @@ if [ "$DO_AUDIO" = "1" ]; then
if command -v nvidia-smi >/dev/null 2>&1 && nvidia-smi -L >/dev/null 2>&1; then
uv_sync_extras+=(--extra gpu)
fi
say "Synthesising narration with Qwen3-TTS (speaker=${TTS_SPEAKER:-ryan}) — one batched call"
say "Synchronising tts/ Python deps"
uv sync --project tts ${uv_sync_extras[@]+"${uv_sync_extras[@]}"} || fail "uv sync failed in video/tts"
uv run --project tts python tts/synth.py || fail "tts/synth.py failed"
if [ ! -s output/audio/index.json ]; then
fail "synth did not produce output/audio/index.json"
fi
for sb in "${STORYBOARDS[@]}"; do
say "Synthesising narration for [$sb] — one batched call"
uv run --project tts python tts/synth.py --storyboard "$sb" \
|| fail "tts/synth.py failed for $sb"
if [ ! -s "output/$sb/audio/index.json" ]; then
fail "synth did not produce output/$sb/audio/index.json"
fi
done
fi
# -- record -------------------------------------------------------------------
say "Recording"
# -- record ------------------------------------------------------------------
# record.ts iterates over storyboards in-process and writes per-storyboard
# recording.webm + narration.json. One Node invocation handles all of them
# so we don't spin up Playwright + GPU/WebGL + auth more than necessary.
say "Recording all storyboards"
APP_URL="$APP_URL" node dist/record.js
if [ ! -s output/recording.webm ]; then
fail "recording.webm missing or empty"
fi
node dist/verify.js output/recording.webm
# -- encode -------------------------------------------------------------------
if [ "$DO_ENCODE" = "1" ]; then
if ! command -v ffmpeg >/dev/null 2>&1; then
fail "ffmpeg not on PATH; rerun with --no-encode if you only need the WebM"
for sb in "${STORYBOARDS[@]}"; do
if [ ! -s "output/$sb/recording.webm" ]; then
fail "[$sb] recording.webm missing or empty"
fi
say "Encoding to MP4"
ffmpeg -y -loglevel warning -i output/recording.webm \
-c:v libx264 -pix_fmt yuv420p -crf 14 -preset fast \
-movflags +faststart \
output/recording.mp4
node dist/verify.js "$sb" "output/$sb/recording.webm"
done
# Poster: a single high-quality JPEG extracted from a representative
# moment in the output timeline. Used as the homepage <video poster=...>,
# which is what the visitor sees before pressing play.
# - -ss AFTER -i = output-side seek, frame-accurate (input-side seek
# would land on the nearest keyframe, drifting back up to ~2s).
# - -update 1 tells ffmpeg the output is a single image, not a sequence.
# - -q:v 2 = high JPEG quality (~95%); poster file is ~120KB at 1080p.
say "Extracting poster frame at ${POSTER_TIME_S}s"
ffmpeg -y -loglevel warning -i output/recording.mp4 -ss "$POSTER_TIME_S" \
-frames:v 1 -update 1 -q:v 2 \
output/poster.jpg
node dist/verify.js output/recording.mp4 output/poster.jpg
# -- encode + mux + publish (per storyboard) ---------------------------------
if [ "$DO_ENCODE" = "1" ] && ! command -v ffmpeg >/dev/null 2>&1; then
fail "ffmpeg not on PATH; rerun with --no-encode if you only need the WebM"
fi
# -- mux narration ------------------------------------------------------------
# Synth already produced per-cue WAVs (in output/audio/); the recorder logged
# each cue's videoTime against the trimmed timeline. Drop the WAVs onto the
# mp4 with one ffmpeg adelay+amix and replace the silent recording in place.
if [ "$DO_ENCODE" = "1" ] && [ "$DO_AUDIO" = "1" ]; then
if [ ! -s output/narration.json ]; then
fail "narration.json missing — recorder did not log cues"
for sb in "${STORYBOARDS[@]}"; do
if [ "$DO_ENCODE" = "1" ]; then
say "[$sb] Encoding to MP4"
ffmpeg -y -loglevel warning -i "output/$sb/recording.webm" \
-c:v libx264 -pix_fmt yuv420p -crf 14 -preset fast \
-movflags +faststart \
"output/$sb/recording.mp4"
# Poster: a single high-quality JPEG extracted from a representative
# moment in the output timeline. Used as the homepage <video poster=...>.
# - -ss AFTER -i = output-side seek, frame-accurate (input-side seek
# would land on the nearest keyframe, drifting back up to ~2s).
# - -update 1 tells ffmpeg the output is a single image, not a sequence.
# - -q:v 2 = high JPEG quality (~95%); poster file is ~120KB at 1080p.
poster_t="$(poster_time_for "$sb")"
say "[$sb] Extracting poster frame at ${poster_t}s"
ffmpeg -y -loglevel warning -i "output/$sb/recording.mp4" -ss "$poster_t" \
-frames:v 1 -update 1 -q:v 2 \
"output/$sb/poster.jpg"
node dist/verify.js "$sb" "output/$sb/recording.mp4" "output/$sb/poster.jpg"
fi
say "Muxing narration into output/recording.mp4"
uv run --project tts python tts/mux.py --replace \
|| fail "tts/mux.py failed"
node dist/verify.js output/recording.mp4
fi
# -- publish to homepage ------------------------------------------------------
# Only publish when we did the encode (otherwise we'd be copying a stale
# mp4 next to a fresh webm). --no-encode skips this whole block.
if [ "$DO_ENCODE" = "1" ]; then
if [ ! -d "$PUBLISH_DIR" ]; then
say "Creating $PUBLISH_DIR"
mkdir -p "$PUBLISH_DIR"
if [ "$DO_ENCODE" = "1" ] && [ "$DO_AUDIO" = "1" ]; then
if [ ! -s "output/$sb/narration.json" ]; then
fail "[$sb] narration.json missing — recorder did not log cues"
fi
say "[$sb] Muxing narration into output/$sb/recording.mp4"
uv run --project tts python tts/mux.py --storyboard "$sb" --replace \
|| fail "tts/mux.py failed for $sb"
node dist/verify.js "$sb" "output/$sb/recording.mp4"
fi
say "Publishing to $PUBLISH_DIR"
cp output/recording.mp4 "$PUBLISH_DIR/recording.mp4"
cp output/poster.jpg "$PUBLISH_DIR/poster.jpg"
node dist/verify.js "$PUBLISH_DIR/recording.mp4" "$PUBLISH_DIR/poster.jpg"
fi
# -- report -------------------------------------------------------------------
# Only publish when we did the encode (otherwise we'd be copying a stale
# mp4 next to a fresh webm). --no-encode skips publish.
if [ "$DO_ENCODE" = "1" ]; then
if [ ! -d "$PUBLISH_DIR" ]; then
say "Creating $PUBLISH_DIR"
mkdir -p "$PUBLISH_DIR"
fi
say "[$sb] Publishing to $PUBLISH_DIR/$sb.{mp4,jpg}"
cp "output/$sb/recording.mp4" "$PUBLISH_DIR/$sb.mp4"
cp "output/$sb/poster.jpg" "$PUBLISH_DIR/$sb.jpg"
node dist/verify.js "$sb" "$PUBLISH_DIR/$sb.mp4" "$PUBLISH_DIR/$sb.jpg"
fi
done
# -- report ------------------------------------------------------------------
say "Done"
if command -v ffprobe >/dev/null 2>&1; then
for f in output/recording.webm output/recording.mp4 output/poster.jpg \
"$PUBLISH_DIR/recording.mp4" "$PUBLISH_DIR/poster.jpg"; do
[ -f "$f" ] || continue
size=$(stat -c '%s' "$f" 2>/dev/null || stat -f '%z' "$f")
case "$f" in
*.mp4|*.webm)
dur=$(ffprobe -v error -show_entries format=duration -of default=noprint_wrappers=1:nokey=1 "$f")
printf ' %s %ss %s bytes\n' "$f" "$(printf '%.2f' "$dur")" "$size"
;;
*)
printf ' %s %s bytes\n' "$f" "$size"
;;
esac
for sb in "${STORYBOARDS[@]}"; do
for f in "output/$sb/recording.webm" "output/$sb/recording.mp4" \
"output/$sb/poster.jpg" \
"$PUBLISH_DIR/$sb.mp4" "$PUBLISH_DIR/$sb.jpg"; do
[ -f "$f" ] || continue
size=$(stat -c '%s' "$f" 2>/dev/null || stat -f '%z' "$f")
case "$f" in
*.mp4|*.webm)
dur=$(ffprobe -v error -show_entries format=duration -of default=noprint_wrappers=1:nokey=1 "$f")
printf ' %s %ss %s bytes\n' "$f" "$(printf '%.2f' "$dur")" "$size"
;;
*)
printf ' %s %s bytes\n' "$f" "$size"
;;
esac
done
done
else
ls -la output/recording.* output/poster.jpg \
"$PUBLISH_DIR/recording.mp4" "$PUBLISH_DIR/poster.jpg" 2>/dev/null || true
fi