THE WORKING RECIPEcontrol_video = the source videodenoise = 1.0+ a night prompt+ your look as the first frame
Every slider below is the SAME moment of the SAME performance, half day and half night. Drag it.
Press play — both halves move together, so you are comparing the light, not the timing.
Does the look hold?
⇄
DAY PLATEL1 — Warm practicals
Chunking — does it degrade?
Chained each chunk starts from the previous chunk's relit frame — smooth, but carries the fade
⇄
DAY PLATECHAINED
Independent every chunk restarts from the look image — strong, but a hard step at each seam
⇄
DAY PLATEINDEPENDENT
chain
step at each seam
normal step
p95
ratio
cast start → end
independent
+52, +56
0.1
2.9
467.51×
35.2 → 62.1
chained
+23, +48
0.3
3.6
168.74×
35.2 → 103.7
The ratio is the whole test: a seam is only visible if it is
much bigger than the step between ordinary neighbouring frames. Under ~5× you will not see it.
L2 · Warm + cool accent warm practicals plus cool blue-violet on the wall · 2 lamps
⇄
DAY PLATEL2
L3 · One lamp + cool top left lamp only, cool light from above · 1 lamp
⇄
DAY PLATEL3
L4 · Cool, no practicals blue and violet only, lamps dark · none
⇄
DAY PLATEL4
Numbers
cast is R−B: the warm/cool axis, and the signal that actually shows a relight.
The day plate sits around +73. drift is the spread across the clip in fifths — a look that
holds is flat. structure is shape correlation against the plate; the motion floor is 0.909.