More FE changes

This commit is contained in:
Andras Schmelczer 2026-05-09 09:43:41 +01:00
parent f114ada255
commit a48eb945e0
48 changed files with 4127 additions and 1751 deletions

View file

@ -10,15 +10,17 @@
# ./render.sh # full pipeline (uses cached auth.json if fresh)
# ./render.sh --fresh-auth # force re-auth even if auth.json exists
# ./render.sh --no-encode # stop at WebM, skip MP4 encode
# ./render.sh --no-audio # skip Qwen3-TTS narration; publish 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
set -euo pipefail
# -- config (override via env) -------------------------------------------------
APP_URL="${APP_URL:-http://host.docker.internal:3001}"
PB_URL="${PB_URL:-http://host.docker.internal:8090}"
API_URL="${API_URL:-http://host.docker.internal:8001}"
export APP_URL="${APP_URL:-http://host.docker.internal:3001}"
export PB_URL="${PB_URL:-http://host.docker.internal:8090}"
export API_URL="${API_URL:-http://host.docker.internal:8001}"
PB_ADMIN_EMAIL="${PB_ADMIN_EMAIL:-admin@propertymap.local}"
PB_ADMIN_PASSWORD="${PB_ADMIN_PASSWORD:-propertymap-dev-2024}"
PB_EMAIL="${PB_EMAIL:-demo-video@local.test}"
@ -34,14 +36,28 @@ PUBLISH_DIR="${PUBLISH_DIR:-../frontend/public/video}"
# 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="${MAX_DURATION_S:-45}"
export MIN_DURATION_S="${MIN_DURATION_S:-10}"
export OUTPUT_FPS="${OUTPUT_FPS:-50}"
FRESH_AUTH="${FORCE_AUTH:-0}"
DO_ENCODE=1
DO_AUDIO=1
for arg in "$@"; do
case "$arg" in
--fresh-auth) FRESH_AUTH=1 ;;
--no-encode) DO_ENCODE=0 ;;
--no-audio) DO_AUDIO=0 ;;
-h|--help)
sed -n '3,18p' "$0"
sed -n '3,20p' "$0"
exit 0 ;;
*) echo "Unknown arg: $arg" >&2; exit 2 ;;
esac
@ -124,12 +140,36 @@ else
say "Reusing existing auth.json"
fi
# -- record -------------------------------------------------------------------
say "Recording"
# -- preflight + synth (Qwen3-TTS) -------------------------------------------
# Synth runs BEFORE recording: one batched generate_custom_voice call across
# all cues so the voice stays consistent. The recorder reads
# output/audio/index.json for measured per-cue durations and sizes each
# cue's wall-clock to fit; --no-audio skips synth and the recorder falls
# back to a worst-case estimate.
mkdir -p output
# Wipe last run's leaking artifacts so the rename step picks up *this* run.
rm -f output/recording.webm output/recording.mp4 output/page@*.webm output/page@*.webm.untrimmed
rm -f output/narration-script.json output/narration.json
# output/audio/ is preserved; tts/synth.py decides whether the cached WAVs
# still match the script and skips generation when they do.
say "Preflight: emitting narration script"
node dist/preflight.js
if [ "$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."
fi
say "Synthesising narration with Qwen3-TTS (speaker=${TTS_SPEAKER:-ryan}) — one batched call"
uv sync --project tts || fail "uv sync failed in video/tts"
uv run --project tts python tts/synth.py || fail "tts/synth.py failed"
if [ ! -s output/audio/index.json ]; then
fail "synth did not produce output/audio/index.json"
fi
fi
# -- record -------------------------------------------------------------------
say "Recording"
APP_URL="$APP_URL" node dist/record.js
if [ ! -s output/recording.webm ]; then
@ -163,6 +203,20 @@ if [ "$DO_ENCODE" = "1" ]; then
node dist/verify.js output/recording.mp4 output/poster.jpg
fi
# -- mux narration ------------------------------------------------------------
# Synth already produced per-cue WAVs (in output/audio/); the recorder logged
# each cue's videoTime against the trimmed timeline. Drop the WAVs onto the
# mp4 with one ffmpeg adelay+amix and replace the silent recording in place.
if [ "$DO_ENCODE" = "1" ] && [ "$DO_AUDIO" = "1" ]; then
if [ ! -s output/narration.json ]; then
fail "narration.json missing — recorder did not log cues"
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.