HAL
LATENCY · making-of · prompts-video

Seedance

Every Seedance 2.5 take in production order: the shot, its inputs by role, the prompt, the window, the price, the verdict and whether it reached the film.

The takes run on fal.ai: Seedance 2.5 reference-to-video (bytedance/seedance-2.5/us/reference-to-video), 720p, 16:9, $0.473 a second. @Image1 is the first frame, @Image2 the face sheet, @Audio1 the shot's window of the vocal stem (kit/windows.py: only the shot's own words, silence around them, 0.7 s before the first). The prompt is the wiki's sung-take template, written by kit/takes.mjs from takes/plan.json, the canon and the timed words; each prompt is saved next to its take as takes/<take>.txt. Lip sync is measured, not guessed: kit/lips.py correlates the gap between his lips, frame by frame, with the loudness of the window.

The pilot · 18 November 2026

TakeShotInputsSecondsPriceVerdictIn the film
S13-b13frame, sheet, window · Seedance 2.5 Flash6$0failed: fal's partner check refused the input pictures as "likenesses of real people"no
S06-a6frame, sheet, window · his voice as the soundtrack13$0 (refused)failed after 5 minutes: "Output audio has sensitive content"no
S13-c13frame, sheet, window · his voice as the soundtrack6$2.84maybe: the soundtrack came back as the window (envelope 0.90, drift under 0.15 s), lips 0.71; but the camera pushed in to an extreme close-up and the teardrop grew into a dripno
S13-a13frame, sheet, window · mute6$2.84keep: locked off like the first frame, on-model, lips 0.67 at 0.04 s (control 0.18)yes, v0
S06-b6frame, sheet, window · mute13$6.15keep: eyes close on the long note, the teardrop shows in the reflection; lips 0.32, 0.41 with the picture 0.42 s lateryes, v0, shifted 0.42 s

What the pilot settled. Flash cannot take his face at all. The full model takes it, but a soundtrack of his singing voice can be refused on the way out, so the takes come back mute: generate_audio: false, the prompt's audio policy rewritten to "@Audio1 is the timing reference for his mouth". The lips follow @Audio1 as well as with sound on (0.67 against 0.71 on the same shot), and the film's sound is the song anyway. Without a soundtrack there is nothing to warp, so each take is placed by its lips: kit/lips.py gives the delay, kit/cut.py applies it when it clearly helps.

The rest · waiting

Fourteen takes, 154 s, $72.84 at list: shots 2, 3, 4, 5, 8, 10, 11, 12, 14, 15, 16, 17 sung, 7 and 9 silent. The run stopped on its first uploads: the fal balance ran out. Nothing was sent, nothing charged.