Motion Graphics Pipeline Review

HyperFrames studies · AdLibrary references · dataset handoffs

Updated 2026-09-18 · static review bundle

Upscaling — resumed on rack 2

Resume snapshot: September 18, 2026, 12:22 PM PDT. Two rack-2 nodes; see the live dashboard for current counts.

2-node cap · rack 2
Processed1,512 / 1,582
Delivered to SeedHost1,500
Still unprocessed70
Awaiting delivery12
Active GPU nodes2
Retry-limited chunks8 in 1 video
Completed outputs and checkpoints are preserved.

The user resumed processing on two rack-2 nodes, with eight GPUs total. The controller and SeedHost transfers are active again. Completed checkpoints are reused; no rack-3 jobs were resumed. The eight retry-limited chunks belong to one video; the worker termination cause remains unresolved.

Remaining reviews

Pending

Small text after upscaling

Compare originals and upscaled outputs for changed letters, unreadable details, flicker, warping and cropped text. Start with the 620 text-sensitive source hints, while checking unflagged clips too.

Poor results need rework and another review. Keep them out of the approved set until they pass. Output resolution alone does not establish text accuracy.

Pending

Actual motion-graphics content

Separate graphics-led videos, substantial animated overlays, mixed videos with usable excerpts, ordinary footage, and uncertain cases. Correct title, logo and infographic labels.

Talking heads, product footage, K-pop performances, subtitles and static branding do not qualify by themselves. Machine labels are candidate evidence, not approval.

Source inventory and context

This run upscaled 1,582 distinct files from the September 10 AdLibrary.com batch: 49.13 source hours, associated with 1,873 downloaded ad records. Of those files, 1,579 were new to the earlier collection. Broad E-commerce and Apps searches favored download volume, so the collection includes reading ads, drama, talking heads and product footage as well as designed graphics.

Platform counts use ad records: 1,868 Instagram and 5 Facebook. The older 581-source / 170-candidate collection is separate. These outputs are candidates for later H3/FastVideo data preparation; this run did not train a model or create approved DFD pairs.

Original resolution

Short edgeFilesShare of this run
360 px84653.5%
404–476 px20.1%
540–576 px110.7%
720 px72345.7%

None of the originals are native 1080p. Production used SeedVR2 7B with a 1080-pixel short-edge target and preserved aspect ratio. Upscaling cannot guarantee recovery of text that was unreadable in the source.

Original orientation

OrientationFiles
Portrait1263
Square318
Landscape1

Existing machine format labels

1,082 classified; 500 unclassified due to processing errors. These are existing sampled-video estimates, not fresh review. Hours below describe entire source videos, not the duration of motion-graphics segments.

Primary video formatFilesFull-source hours
Text story or reading ad61416.449
Processing error50013.516
Live action story or drama20212.499
Live action product demo651.149
Talking head ugc or testimonial612.110
Narrative animation or vfx502.444
Product broll or montage460.471
Slideshow or static graphics240.160
Performance music or dance60.059
Other or mixed50.066
Animated infographic or explainer50.185
App or software demo20.009
Motion graphics10.005
Gameplay10.006

On a separate graphics-presence axis, the classified files include 86 graphics-led, 208 substantial integration, 626 incidental text/logo and 162 none. Do not add these to the primary-type counts or treat them as accepted clips.

Last estimate before stopping

Historical only: at approximately 8:50 AM PDT on September 18, remaining processing was estimated at 8–10 hours and final delivery at 10–14 hours, excluding blocked-video recovery. Those estimates assumed six nodes before the pause and no longer apply. The resumed two-node run needs a new estimate from measured throughput.

The SeedHost run folder contains UPSCALING_HANDOFF.md, the 1,582-row UPSCALING_INVENTORY.csv, outputs and receipts. Internal server locations and private transfer configuration are not published here.

Download this public status snapshot · View the small model comparison sample · Live cluster allocation page

No visual/audio QA or content review was run for this update. Text-quality approval and motion-graphics approval remain separate from processed/delivered counts.