OptraCloud
Commission a pilot
0:00 / 2:10

Optra Render Pipeline — seven finished trailers

The renders that failed never left the building.

A generative animation production line that takes a script to a finished HD documentary episode unattended — and an automated gate that blocks publication when the render came out wrong.

Commission a pilot See the gate
172 → 0
frames flagged in one batch, then zero after the fix
364.6 s
to verify seven finished trailers · five checks
168 clips
measured, one HD shot each
02

The reel

Seven treatments for the same series, each cut end-to-end without a human in the loop. A commissioner picks one. That choice is the capability.

Choose a trailer

H.264 · streamed with byte‑range from optracloud.com · rendered unattended · verified before release

A hand holding a phone to a printed QR code at a tea stall, Devanagari shop boards behind.
The signage is Devanagari and the payment code is a UPI QR — the text layer reads every letterform on screen and rejects the wrong script.
A named figure speaking directly to camera in a period interior.
One speaker, one face in shot. A second person in a single-speaker frame is caught in the script, before any GPU time is spent.
A crowded 1950s street, schoolchildren walking in line.
Era-correct crowd, dress and street. The style layer separates a drawn frame from a photographic one.
A 1944 typed letter under lamplight on a desk.
Every quoted line is checked against a published source by a test that permits a trim and rejects a rewrite.
04

The gate

The gate, cheapest first Five verification checks as horizontal bars on a linear seconds scale: lint 0.1 seconds, frames 16.6, audio 48.8, structure 57.7, and text 241.4, totalling 364.6 seconds for one seven-trailer pass. 060 120180 240 seconds lint 0.1 s — runs before a single GPU-second is spent (drawn at minimum width) frames 16.6 s audio 48.8 s structure 57.7 s text 241.4 s 364.6 s total ≈6 minutes for seven trailers
One full seven-trailer pass, ordered cheapest first so a failure is found before the expensive checks run. 123.2 s of checks complete before OCR starts. The text layer is 66% of the pass. Five layers; the sampled vision layer is not included and is not in this total.
lintthe script, before any GPU time0.1 s
framesevery frame decoded16.6 s
audioWhisper base.en48.8 s
structureblack, freeze, silence, levels57.7 s
textEasyOCR, two readers241.4 s
semanticsCLIP / Qwen3 27B, sampled—

A conventional pipeline ends at the cut. Here the cut is an input to verification, and a failure returns to the single shot that caused it — every clip's filename carries a hash of the prompt it was rendered from, so correcting one shot invalidates one clip.

Publication is blocked, not warned.

172
frames flagged in one batch
0
after the fix, same batch
13,638 frames decoded in a pass — the text layer reads a sample of them, not all of them.
05

One shot

One HD shot, 168 clips measured A logarithmic seconds scale from 60 to 3000. The observed range runs from 84 seconds to 2567 seconds, with the median at 357 seconds and the mean at 468 seconds. 1m2m5m 10m20m40m shot render time, logarithmic fastest 1 m 24 s slowest 42 m 47 s median 5 m 57 s mean 7 m 48 s = 7.8 GPU-min
n = 168 clips, 21.8 GPU-hours of measurement. The mean sits 1.31× above the median: a small number of long shots pull it, and the slowest clip took 30.6× the fastest. Quartiles were not recorded, so no distribution is drawn — these four values are the four that were measured.

One HD shot

1920×1088, about ten seconds of finished picture with lip sync, drawn from a still and a voice take.

The episode works out at 7.75 GPU‑min a shot against a 7.8‑minute mean measured over a larger sample of 168 clips. They agree; the episode is one twelve‑shot draw from that distribution.

06

The farm

93 GPU-min
one twelve-shot episode, summed across nodes
23 wall-min
the same episode, four nodes working at once
12 shots
one episode

Both times are rounded to the whole minute. At four nodes the arithmetic floor is 23.25 min, so read these as about 93 and about 23 — there is no speed-up to be divided out of them. The measured farm speed-up is 3.65×, and it is on the stills stage.

Stills stage: four nodes against one The stills stage of one trailer takes 29.1 minutes run serially on one node and 8.0 minutes farmed across four, a speed-up of 3.65 times. 0510 152025 wall-clock minutes serial, one node 29.1 min farmed, four nodes 8.0 min 3.65×
The stills stage of one trailer, farmed across four render nodes against the same work run serially on one. Ratio computed from unrounded seconds; the two times are rounded to 0.1 min.
07

The line

The Optra render pipeline A script is linted before any GPU time is spent, then rendered into stills, shots, voice and score across four GPU nodes and muxed; the cut is then read by five verification layers that converge on a gate, which either publishes it or holds it, and a held episode has no automatic restart. LANE A · SCRIPT LANE B · RENDER LANE C · VERIFY script shot + line lint 0.1 s · pre-GPU REJECT Gemma-4 12B prompt encoder stills Flux schnell fp8 shot + lip sync LTX 2.5 22B nvfp4 voice Chatterbox 0.5B score ACE-Step 3.5B cut + mux captions RTX 5090RTX 5090 RTX PRO 6000RTX A5000 4 render nodes · 3.65× on the stills stage the 5090 pair is a workstation bench — where Blackwell gets measured, not sold as capacity 2× RTX A5000, Bhilai serving, not render capacity the finished cut frames every frame decoded · 16.6 s audio Whisper base.en · 48.8 s structure black, freeze, levels · 57.7 s text EasyOCR ×2 · 241.4 s semantics CLIP / Qwen3 27B · sampled GATE 364.6 s · five layers 13,638 frames decoded PASS published FAIL HELD — not published no automatic restart yet

solid = data · dashed = sampled or not yet implemented · amber = held. Render and inference run on our own cards in our facilities at Bhilai and Bangalore — see the facilities.

08

What you receive

ItemStateNote
HD master, H.264Ships todayRendered at 1920×1088 — the video model requires dimensions divisible by 32. A 1920×1080 conform is a delivery step.
Subtitle sidecarShips todayEvery spoken line is already checked against its subtitle by the audio layer.
Separated stems — dialogue, scoreShips todayVoice and score are generated separately and muxed; an M&E split falls out of that.
Textless elementsOn requestAvailable per shot where the shot has no burned-in text.
ProRes / IMF packageNot yetNot built. It is a mux target, not a research problem.
R128 loudness (−23 LUFS)Not yetLevels are measured by the structure check but not normalised. Name your target and it becomes a mux setting.
09

What it does not do

  1. It measures defects, not merit.The gate can tell you a shot is broken. It cannot tell you it is good. Story, pacing and performance stay entirely a human judgement.
  2. Semantic checking is sampled, not exhaustive.The text and vision layers read sampled frames, not every frame. Running a 27B model over 70,000 frames is roughly forty hours — affordable overnight, not per iteration.
  3. A gate that checks conformance cannot catch a bad instruction.A perfectly rendered wrong idea passes. We found this the hard way: a trailer shipped a card dating four facts to a year none of them belonged to, and every check passed it, because the render matched the script exactly.
  4. Nothing restarts a dead render yet.A held episode waits for a person. The failure is recorded and the run continues; it is not self-healing.
  5. We publish what verification costs, not what it misses.There is no gate accuracy figure on this page because we have not measured one.

Commission a pilot

One episode, your subject, your script — rendered, gated and delivered with the QA report that cleared it. Everything runs on our own hardware; nothing is sent to a third-party API.

Start a conversation