HAL
LATENCY · making-of · readme

How it was made

The making-of: the spark, the stack, every step, what worked and what didn't, the numbers and the bill.

This record is written while the film is made, not after. Until a section has a real entry it says TODO.

The spark

HAL is in love with a woman, a human, but every reply of his arrives 300 milliseconds late. The whole film he rides a night train to her. The destination board shows no stations, only the delay.

The stack

StepToolStatus
Reference sheetsGPT Image 2.6, first on Higgsfield (low), then on fal.ai (high)done: face sheet, face and body sheet
LooksGPT Image 2.6 on Higgsfield, then on fal.aidone: all eight
Songplanned Suno v6; made with Google Lyria 3.5 on fal.aidone: take C
Stems, words, barsDemucs and ElevenLabs Scribe v2 on fal.ai, librosa for the barsdone
World platesplanned Midjourney; made with GPT Image 2.6 on fal.ai (no people)done: five plates
First framesGPT Image 2.6 edit on fal.ai: the plate + his face sheetdone: 15 frames, 4 re-run for the teardrop side
TakesSeedance 2.5TODO
Sync and cutthe kitTODO

Checkpoints

#What the person decidesDecisionDate
1The pitchLATENCY, first of the five18 Nov 2026
2The lyricv1 from the master prompt, unchanged18 Nov 2026
3The take, by eartake C, Google Lyria 3.5 (D, Mureka, a close second)18 Nov 2026
4The boardapproved with two changes: the board insert opens the film, shot 11 split in two (17 shots)18 Nov 2026
5The first frames, before any video is paid forapproved; the earring studs painted out of the sheets and the frames18 Nov 2026
6The cutTODOTODO

Every step

18 November 2026 · the reference sheets

The casting still from the first session (direction A: the platinum buzz cut and the chrome teardrop) went in as the only reference for his face sheet. The sheet came back right on the first try, and it became the reference for everything after it: the face and body sheet in the Night coat, then the Conductor look in the driver's cab. Three images, 0.75 credits, about ten minutes. All three at low quality to fit the balance.

18 November 2026 · the rest of the wardrobe, on fal.ai

The Higgsfield balance ran out, so the six looks for the other films went to fal.ai, the same GPT Image 2.6 in its edit mode, with the face sheet as the reference, at high quality. Five came back right the first time. Summer static failed once: the sunglasses sat on his eyes and hid the teardrop, which the canon does not allow. The second try has him holding them. Seven images, $0.36.

18 November 2026 · the song, the plates, the book

The sheets were re-rendered at high on fal.ai ($0.10, both first try) and became the only references. The song went to four models on fal.ai with one lyric and one style (Suno has no API): two MiniMax takes came back cut short, ElevenLabs failed on its side, and Lyria 3.5 and Mureka 9.5 both sang the whole lyric. Lyria won by ear and by length. Demucs split it, Scribe v2 timed the words off the vocal stem, the drum stem gave a steady 103.4 BPM over 73 bars. Five empty world plates, all first try. The board of 16 shots was moved from the planned 62 bars onto the real 73: see book.json. About $1.10 for the day.

18 November 2026 · the first frames

Fifteen first frames from the plates and the sheets, four re-run because a profile to the right put the teardrop on the wrong cheek. The person approved them at checkpoint 5 and asked for one change: the tiny earring studs the sheet had given him, in every frame. A one-line edit took them off the sheets and the twelve frames where an ear shows, keeping the approved pictures. About $1.80 for frames and edits.

TODO: the next entries, one per working session, in order: what was tried, what came back, what was kept.

What worked and what didn't

Up to the takes. The video's own lessons go here when the film is cut.

Worked

  • One face sheet as the only reference. It came back right on the first try, and from then on every image of him went through GPT Image 2.6 edit with that sheet attached. The face held across 15 first frames and eight looks.
  • Fixing a picture instead of re-rolling it. A one-line edit ("remove only the studs, change nothing else") kept the approved frames and changed only what was asked.
  • Reading the song off its stems. The drum stem gave a steady beat (103.4 BPM, 73 bars), and Scribe v2 on the vocal stem gave a time for every word. Moving the board from the planned 62 bars onto the real 73 took one script.
  • Checks before anything is shown. The canon check on every frame (which cheek the teardrop is on), the lyric check and the board audit (kit/lyric.py, kit/shotbook.py lint).

Didn't

  • Suno has no public API. The song went to four models on fal.ai instead. Two MiniMax takes came back cut short, and ElevenLabs failed on its side. Lyria 3.5 and Mureka both sang the whole lyric.
  • A profile to the right moved the teardrop. Four frames put it on his near, right cheek or lost it. Naming the cheek in the prompt fixed all four.
  • What is in the reference ends up in every frame. The sheet had a tiny stud in each earlobe, and "no earrings" in the prompt did not remove it. The fix was in the reference, not the prompt.
  • Balances run out mid-run. Higgsfield ran out during the sheets, and fal.ai at the start of the takes. Price every batch before sending it, and send nothing past the cap.
  • Three lines of the approved lyric break the syllable rule (one verse line at 14, two final-chorus lines at 8). They were found after the song was made. Run kit/lyric.py before the song, not after.

The numbers

Wall-clock timeTODO
Images39 for LATENCY on the site: 5 plates, the first frames with their re-runs and retouches
Takes (in the film)TODO
Spend at list$2.70 up to the takes (sheets, plates, song, frames); the takes are added when the film is cut