← Blog
✎blogbericht

Fifty Stills and Four Clips for One Window Cleaner, and 39 Pictures Finished

A steel film reel on a dark slate bench holding a short strip of exposed frames, cold teal light with a thin amber rim along one flange, empty space above the reel.

One batch rendered 50 stills at 10 to 14 seconds each, and four 288 by 512 clips took 130 to 270 seconds of wall time, but only 39 finished pictures reached the output folder and no clip has been judged.

Part 4 of 4 in Image models in a real pipeline. Previous: Krea Left the Pipeline, and One Post Still Says Otherwise · Next: none yet

For one window-cleaning site we queued 50 stills in one batch, and four clips: three of about 10 seconds and one of 15. The stills took 10 to 14 seconds each. The clips took 130 to 270 seconds each. The stills batch did not finish: 39 pictures reached the output folder, and 11 never did.

The stills

The batch file lists 50 entries, all Z-Image Turbo, rendered at 1536 by 864 pixels with nine steps and a shift of 3. The first five prompts describe the chalet, the squeegee on the balcony, water on glass, and an office facade. The prompts ask for sharp focus and true colour.

The first pass rendered all 50 pictures. The last six timings in its log run from 10.3 to 13.8 seconds each.

The second pass upscales each raw picture. It completed 20 of them, then stopped on picture 21, whose raw PNG was never written. The folder holds only the prompt text for that picture. A third pass over the remaining pictures logged 23 more successes. That makes 43 successes across two passes, but the output folder holds 39 JPEGs. Four successes are not accounted for, and 11 of the 50 pictures have no final file at all.

We have not explained the four. The logs also print a raw shape of 5760 by 4608 pixels for upscaled files, against the 1536 by 864 request. The log does not say which shape is the raw render and which is the upscale, so we cannot yet say what size the files are.

The clips

Three clips were rendered at 241 frames, and one at 361 frames. The graph files record 288 by 512 pixels at 24 frames per second, with two sigma schedules, one of eight steps and one of three. That is a two-stage graph.

Frames at 24 per second give the video length: 241 frames is 10.04 seconds, and 361 frames is 15.04 seconds. The run logs report graph wall times of 145.6, 129.5 and 130.1 seconds for the 241-frame clips, and 268.6 seconds for the 361-frame clip.

Dividing wall time by video length gives about 13 to 14.5 seconds of compute per video second for the short clips, and about 18 for the long one. That is the real-time factor, and it is far above the 0.83 we reached on a smaller live profile: 384 by 224 pixels, three steps, one stage. Those are different jobs, so the comparison shows the gap, not a regression.

The same box took 69 to 81 seconds per shot for a 121-frame, 576 by 1024, two-stage ad in the winter-dog work. That is a different shape again.

Two exit codes after PASS

Both run logs print PASS for every graph, then a line saying the wrapper job failed with exit code 1. We do not yet know whether the wrapper or the graph is at fault. The output files exist, so the graph ran. We have not traced the wrapper.

What we did not measure

No clip has been judged by eye, and no judgment of the clips is recorded. We have not probed the three MP4 files for their actual frame counts or durations. The stills have not been scored against the prompts, and the 39 that exist have not been looked at for this post.

How to prove this wrong

Count the JPEGs in the output folder. If it holds 50, the 11-picture gap is wrong. Probe each clip with a media tool and read the frame count and duration. The three 241-frame clips should report 241 frames and about 10 seconds, and the 361-frame clip about 15 seconds. Then rerun the 361-frame graph twice. If the warm wall time is far from 268 seconds, the single timing was an outlier.

The open question is the one that changes the plan: is a 50-picture batch worth finishing if only 39 pictures are needed for the site, and which of the clips is worth judging first?

Provenance

  • /home/mario/Projects/web-apps/amarbaroHQ/.work/fensterglanz/stills.json (50 entries, first five settings).
  • /home/mario/Projects/web-apps/amarbaroHQ/.work/fensterglanz/stills/pass1.log (50 successes, last six timings).
  • /home/mario/Projects/web-apps/amarbaroHQ/.work/fensterglanz/stills/pass2.log (20 successes, then the missing-file stop at picture 21).
  • /home/mario/Projects/web-apps/amarbaroHQ/.work/fensterglanz/stills/pass2b.log (23 successes from the remaining pictures).
  • /home/mario/Projects/web-apps/amarbaroHQ/.work/fensterglanz/stills/out/ (39 JPEG files).
  • /home/mario/Projects/web-apps/amarbaroHQ/.work/fensterglanz/video/long/v1long.json, v2long.json, v3long.json, v1long15.json (frame counts, 24 frames per second, 288 by 512, sigma schedules).
  • /home/mario/Projects/web-apps/amarbaroHQ/.work/fensterglanz/video/long/run.log and run2.log (graph wall times and the exit-code lines).
  • /home/mario/Projects/AssetsHQ/docs/lanes/video-fast-report.md and /home/mario/Brain/AssetsHQ/2026-11-17-video-fast.md (the 0.83 real-time factor and its 384 by 224 shape).
  • /home/mario/Brain/AssetsHQ/2026-11-19-winter-dog-ltx-commercial.md, line 17 (the 69 to 81 second shot times).

Reacties

Nog geen reacties.

Inloggen om te reageren.