ratio must be "adaptive" for frame-pinned clips — current guidance breaks seams above 480p

[domtisdell]

ratio must be adaptive for frame-pinned clips — the current guidance breaks seams above 480p

Summary. The skill tells you to always pass ratio explicitly. For the chained
first/last-frame technique the skill is built around, that instruction is wrong: it
makes Seedance refit the pinned frame instead of reproducing it, and the error
compounds down the chain.

From SKILL.md (Monid backend, and repeated in Gotchas):

Pass ratio explicitly (16:9, or 9:16 for the mobile chain) — the adaptive
default follows the input image's aspect instead.

Monid clip wrong aspect → the ratio default is adaptive and follows the input
image… Pass --ratio … explicitly on every chained clip.

That is right for text-to-video. It is the opposite of what frame-pinned generation
needs.

Measured

Same leg, same prompt, same frames; only ratio changed. Frame 0 of the output
compared against the pinned first_frame:

resolution	ratio	frame-0 match
480p	16:9	33.7 dB
720p	16:9	16.0 dB — a pure ~2% zoom, no translation
1080p	16:9	16.1 dB — same 2% zoom
1080p	adaptive	37.1 dB

Confirmed across a full six-station film: mean seam 36.2 dB at 1080p with
adaptive, no correction applied, no re-rolls. The same chain at 480p measured
33.7 dB. Provider: Monid bytedance /v1/video/seedance-2.0, 2026-08.

Why 480p looks like it works

At 480p the model returns 864×496 — not 16:9. It is ignoring the requested ratio
at that tier, so the frame passes through untouched. That makes 480p appear to be the
only reliable resolution and sends you looking for a resolution bug. It isn't one.

Dead ends, so nobody repeats them

Diagnosing this took a good many wasted re-renders. Three encoder-side corrections
were tried and all scored worse than an uncorrected clean render:

• translation-only realignment — wrong transform; the error is scale, not shift
• shrink-and-pad — 27.0 dB mean, plus a cream border that grows down the chain
• crop-and-zoom — 25.8 dB mean

The 2% error compounds ~10% across six legs, so any cumulative correction also
progressively resamples the footage — losing the sharpness the higher resolution was
for. The fix has to be at render time.

Third-party docs agree, e.g. Runware: "the output size comes from the first frame";
Pixazo: "for first/last-frame generation, ratio must be adaptive". Monid's own
Seedance 2.5 notes say editing/extension require adaptive.

Suggested change

Qualify the existing rule rather than replace it:

Pass ratio explicitly for text-to-video. For frame-pinned clips (anything with
first_frame/last_frame — i.e. every chained leg and connector) pass
ratio: "adaptive": a forced ratio makes the model refit the pinned frame (~2% zoom)
and the error compounds along the chain. 480p masks this because it ignores the
ratio and returns 864×496.

Worth a line in Gotchas too, since the symptom presents as "seams only lock at 480p".

Happy to open a PR if useful.