Vertical Video Dataset Review

9:16 source screening · subtitle repair · dance pilot · GB200 production

Updated September 1, 2026 · live progress snapshot
20.0 / 50 hours Completed delivery runtime
1,549completed deliveries
86.2htotal accepted pool
2 jobsactive cluster compute
0production failures

Source Inventory

Only public portrait, non-AI-generated Shorts without persistent source branding are promoted.

Production and source expansion active

DramaBox

Original production source. Free/public inventory resolved and watermark-filtered.

302 complete
1,082raw
302accepted
6.2hruntime

GoodShort

High branding rate; conservative automatic acceptance retained a small clean subset.

11 complete
324raw
11accepted
2mruntime

ReelShort

Diskless source streaming into NFS, followed by cluster-native OCR and mask preparation.

1,190 / 1,205 complete · 15 remaining
2,315raw
1,205accepted
13.4hruntime

ShortMax

Strong sample yield with public 1080x1920 Shorts and minimal persistent branding.

Full screening candidate
187raw
28/30sample accepted
37.5smean

NetShort

Strong sample yield with public portrait Shorts and low watermark rejection.

46 complete
138raw
46production
34mruntime

DramaWave

Large inventory, but persistent corner branding materially lowers usable yield.

138 / 255 prepared · 0 complete · 6 unavailable
613raw
261accepted
4.8hruntime

MiniShort

Official short-drama inventory screened for public portrait media and source branding.

48 / 48 prepared · 2 complete · 5 unavailable
65raw
53accepted
1.6hruntime

RapidTV

Longer official vertical episodes prioritized for closing the remaining runtime gap.

30-item pilot screening active
139inventoried
30pilot
activescreening

MoboReels

Mixed live-action and AI-labeled catalog. Excluded from accepted runtime until item-level origin verification.

AI-origin review required
1,716inventoried
20public portrait pilot
activescreening

FlickReels

Mixed-origin catalog. Caption removal is blocked until each item passes visual non-AI review.

Pilot rejected as AI-generated
3,121inventoried
29public portrait pilot
activescreening

Delivered Samples

Short excerpts for rapid inspection of final caption-removal delivery encoding.

H.264 CRF16 · audio retained
uB-5XR26to0

DramaBox production sample

DramaBox15sdelivered

Pipeline

Outro filtering, OCR mask generation, ProPainter temporal prior, DiffuEraser 2-step repair, native masked composite, and CRF16 delivery.

Format720x1280 · 30 fps · H.264 + AAC
StatusComplete with receipt
ReviewSelected excerpt only; full-corpus manual QA not run.
1QanjafOKeY

GoodShort production sample

GoodShort15sdelivered

Selection rule

Accepted only after public-access, portrait-format, persistent-logo, and branded-outro screening.

Format720x1280 · 30 fps · H.264 + AAC
StatusComplete with receipt
Source yield11 of 324 Shorts auto-approved.

Projected Throughput

Observed rates from the active rack-2 render and EC2 screening workloads.

Estimate, not delivery commitment
Watermark screening~109 videos/hour across two T4 OCR workers.
Current screen ETAApproximately 1.5–2 hours for the selected RapidTV and MoboReels batch.
Cluster rendering~22 videos/hour observed on one 4xGB200 job.
Overnight outputApproximately 3–6 additional rendered hours after high-resolution staging begins.
New screened pool187 accepted videos totaling approximately 59.94 hours.
Total accepted poolApproximately 86.2 hours including completed and newly screened media.
50-hour delivery ETAApproximately 2–3 days at the two-job / 8-GPU cap.
EC2 baselineFour g7e.4xlarge workers in Oregon at $3.998/hour each.
EC2 cost to 50hApproximately 13 wall hours and $207 aggregate on-demand compute.
EC2 cost to 100hApproximately 40 wall hours and $643 aggregate on-demand compute.
Cost boundaryCompute-only benchmark estimate; excludes EBS, snapshots, transfer, preparation, and capacity delays. Current Oregon G7e Spot placement score is 1/10.

Operations

Restartable state and durable storage locations.

Compute policyMaximum two concurrent 4xGB200 Jobs, pinned to rack 2; 8-hour deadline per Job.
Local storageAuthenticated downloads stream directly to NFS. MiniShort and DramaWave are paused after repeated unavailable-source retries.
Completed roots20260829-dramabox-scale-b001 · 20260830-dramabox-scale-b002 · 20260830-goodshort-auto · 20260901-netshort-auto
Active root/mnt/nfs/vlm-tzhou/h3-subtitle-repair/20260830-reelshort-auto
EC2 fallbackValidated persistent G7e.2xlarge runtime on encrypted EBS; no EC2 compute currently running.
CleanupQP0 intermediates and completed Kubernetes Jobs are deleted; sources, masks, priors, deliveries, logs, and receipts remain.