Section 02 pour — round 2

Candidates only. Nothing here is on the live homepage.

What was asked for: one solid stream of two mixed colours that stays a single tight column the whole way down, and separates into two mounds only at the moment it lands — with the two mounds touching each other. Smoother and prettier than the shipped clip, which fanned out high and left a gap between the piles.

Video candidates

PV-02
FLUX 3 · 6.0s · 1080p
Strongest. Column holds one width the whole drop and parts only at the floor. Mound-growth correlation +0.95 — the piles are visibly built BY the powder landing, which is the thing that was wrong before.
PV-03
FLUX 3 · 8.0s · 1080p
Same behaviour, slower and calmer, thinner ribbon. Correlation +0.58. Runs 8s, so it would be sped up to ~6s rather than trimmed.
PV-01
MiniMax H3 · 6.6s · 2K
Reject — shown for comparison. The column fans into a wide cone from a third of the way down: the exact 'spreads apart back to the base' problem. Correlation +0.59, and the capsule never fully empties.

Conditioning frames these were built from

START-A
Start frame A
Used as frame 0. One tight two-tone column, splitting only at the floor into two touching mounds.
START-B
Start frame B
Alternate. Tighter column, but the two piles merge into one — not used.
END-A
End frame A
Used as the final frame. Capsule spent, air clear, cap resting on the floor, both mounds built and still touching.

Both keyframes are pinned, so the beginning and the end are composed rather than hoped for — that is what previously kept the mounds from building and the cap from settling. Everything is a full generation; nothing is composited.