fix: captions you can read on a light card, and headlines in capitals
openreels-deploy
5d7a3f5
feat: carry the footage credits to the video description
openreels-deploy
da29a20
perf: download the smallest clip that still fills the frame
openreels-deploy
6973218
fix: a render that times out is tried again, and the budget is bigger
openreels-deploy
ef06b20
feat: give every scene the same look
openreels-deploy
381ef9f
feat: real footage instead of slides, and cards that carry a headline
openreels-deploy
58a35f9
fix: a named model is a named model, and the opening has to be specific
openreels-deploy
d788a73
feat: show where each fact came from
openreels-deploy
849bf4c
The research stage now keeps the pages it read, and the job page lists
them. A fact can be checked without going back to the search.
feat: cut the footage down before the renderer sees it, and write for attention
openreels-deploy
deca22e
A stock library hands over a four-minute 4K file. The renderer pays for
every frame of it, so one scene cost minutes of CPU and megabytes of
upload. Clips are now trimmed to a scene length, scaled to 720p and
stripped of their audio, which the narration covers anyway.
The script rules now care about attention: the first line states the
tension, every scene changes something and adds new information, the turn
is early, and the last line asks for something.
fix: judge the footage with an answer the gateway can give
openreels-deploy
c70f33f
The verification asked the model for a schema object. The gateway cannot
produce one, so every frame came back unverified and the first candidate
was used on trust - the wrong subject in the finished video.
The judge is now asked for plain JSON and the answer is parsed here, which
works on any OpenAI-compatible endpoint. A frame the judge could not look
at is kept as a last resort instead of ending the search, so the other two
stock libraries still get their turn.
fix: research in code, and keep the credit with the footage
openreels-deploy
8f6a281
The research stage asked the model to call a search tool. A gateway behind
an OpenAI-compatible endpoint cannot be relied on to do that, and when it
did not, the stage was skipped and the script was left with one fact.
The search is now made by the pipeline and the model turns what came back
into a brief. Tavily is used when a key is set, Wikipedia when it is not.
A search that fails no longer ends the stage.
Every Wikimedia asset now writes its licence, author and source URL into
credits.json beside the finished short.
feat: a better narrator, every stock library, and research that runs
openreels-deploy
49489bc
Three things were quietly costing quality.
The voice was Piper. Kokoro is a local neural narrator: real intonation,
steadier phrasing, and it reads like a person rather than a list of facts.
It is now the default voice, baked into the image with its model and voice
files, with Piper behind it as a fallback so a missing file cannot end a
render. The build proves the narrator speaks before the image ships.
The footage was thin. The judge asked whether a frame matched the request
exactly, so a parked vintage racing car was rejected for a scene about a
vintage racing car, and the scene became a text card. The judge now judges
the subject, and a near miss is used instead of thrown away. Wikimedia
Commons joins Pexels and Pixabay: it needs no key, it is free-licensed, and
its motorsport archive is deep.
The research was skipped whenever a caller supplied a fact, which left the
product with one fact and nothing to say about it. The research stage runs
again, with the caller's fact as part …
feat: master every voice to the same channel, and write the script for the ear
openreels-deploy
2503d13
The voiceover was whatever the engine returned: different loudness and
tone per provider, symbols read aloud, dashes read as words, and a
medium-quality local voice. It now goes through one chain regardless of
engine.
Before synthesis the script is rewritten for a narrator: units and
symbols spelled out, abbreviations expanded, dashes turned into commas,
and sentences over 28 words broken at a pause the writer already wrote.
After synthesis the audio is high-passed, de-essed, and normalised in two
passes to -14 LUFS with a -1.5 dBTP ceiling at 192 kbps. Mastering is
gain and tone only, so word timings and captions stay exact; a missing
ffmpeg leaves the original audio rather than failing the render.
Piper now defaults to en_US-ryan-high with a 0.35s sentence pause, OpenAI
to onyx with documentary instructions, Gemini to Charon, and ElevenLabs
to steadier settings.
Also fixes the sound cues: the deploy tree had them nested one directory
deep, so the pipeline reported them missing and…
fix: check the model cache where transformers.js actually writes it
openreels-deploy
f554a33
fix: never pass a negative volume, and bake the alignment model into the image
openreels-deploy
bbef654
The ducking window is symmetric around each word, so a word longer than
twice the ramp produced a duck above 1 and a negative music volume, which
Remotion rejects - the first render with the new mix failed at assembly.
The duck is clamped now and the volume maths is unit tested across word
shapes. Sound cues are only mounted when their files were copied, so a
missing asset silences the cue instead of failing the render.
The alignment model is fetched at build time. transformers.js caches to
./.cache and a fresh container was spending minutes downloading it inside
a scheduled job.
fix: extract the piper binary where it is expected and prove it runs at build time
openreels-deploy
bea7c89
The release tarball has a top-level piper/ directory, so extracting it
into /opt/piper put the binary at /opt/piper/piper/piper and the chmod
landed on the directory. Every scheduled run since the narrator was added
died at the voice stage with EACCES. Strip the wrapper directory, set the
mode explicitly, and synthesise a test line during the build so a broken
engine fails the build instead of a 00:00 schedule.
Also: the pipeline now writes a poster frame for each finished short, the
home page plate and gallery cards use it, and the visual styles sit in a
wrapping grid instead of a strip that cut the last card in half.
feat: advertise the local narrator and fail closed without a key
openreels-deploy
23a4be0
10 days ago
feat: lock the server behind a key, faster narration, real sound design
openreels-deploy
2b4ed52
- the API schema and docs are no longer public; only health stays open
- OPENREELS_AUTH_REQUIRED makes the server fail closed without a key
- the web app opens on a sign-in gate
- CRF 23, an optional quality review, and Piper for the voiceover
- the text card is a documentary lower third, the music ducks under the
narrator, and a whoosh marks each cut
perf(pipeline): skip the research stage when a direction is supplied
openreels-deploy
7a02506
fix(llm): bound how long one model may spend retrying a rate limit
openreels-deploy
b66f76b
fix(worker): treat every active job as orphaned at startup
openreels-deploy
02f86d9
fix(worker): poll for orphaned jobs faster than BullMQ does
openreels-deploy
42f020f
fix(worker): fail jobs a server restart orphaned
openreels-deploy
61c2917
The lock is held for ten minutes at a time because a render takes
longer. A container replacement mid-render leaves the job in the
active set until that lock expires, blocking the single worker slot
while BullMQ re-runs it for another full pipeline.
On startup the worker now fails any active job whose lock is gone.
fix(render): pass the delayRender timeout to the renderer
openreels-deploy
525a334
The previous fix set the Remotion CLI config, which the CLI reads and
forwards but renderMedia never reads. A programmatic render kept the
renderer's 30s default (28s delayRender budget) and the assembly stage
failed with "not cleared after 28000ms".
The timeout is now passed to renderMedia/selectComposition directly
(120s default). Concurrency is capped at 1 and the OffthreadVideo frame
cache pinned to 1GB.
feat(llm): model fallback chain + scoped structured-output fallback
openreels-deploy
60e647e
- Classify structured-output errors: only capability errors retry as plain
JSON; timeouts, 401/403/429/5xx, cancellation and finishReason
"length" propagate so a backup model can take over.
- Add FallbackLLM (ordered chain, wall-clock budget, abort on cancel),
wired through the provider factory, job creation, OpenAPI and worker.
fix(llm): abort requests that hang past 180s
openreels-deploy
27ebb51
A silent gateway connection left the research stage 'running' for over
20 minutes. Every generateText call now carries AbortSignal.timeout so the
stage fails instead of hanging the job; OPENREELS_LLM_TIMEOUT_MS overrides
the ceiling.
fix: survive LLM rate limits and concurrent meta writes
openreels-deploy
4c6872c
- Retry gateway 429s with exponential backoff + jitter (honours Retry-After).
The Creative Director retried three times with no delay, so a rate-limited
window failed the job instantly with 'Too Many Requests'.
- Give each meta.json temp file a unique name. The API server and worker both
write meta, and the shared .meta.tmp let one writer rename the other's file
away: ENOENT .meta.tmp -> meta.json.
- Regression tests for both (572 passing).
11 days ago
fix(render): raise the Remotion delayRender timeout in the renderer path
openreels-deploy
f83b02b
remotion.config.ts is only read by the Remotion CLI, so the programmatic
renderMedia() call still used the 28s default and every assembly failed on
the Space with 'A delayRender() ... was called but not cleared after 28000ms'.
The timeout is now set through the Remotion client config.
OpenReels on Hugging Face Spaces: automation REST API, no-AI-assets mode, stock-clip cap and render timeout fixes