Section 02 — every pour animation, in order

All 25 clips generated for the capsule-opens-and-spills shot, newest family last. Each one carries its ID burned into the corner, so any take can be called out by name. Previews are downscaled to 540×960 for loading; the untouched masters stay on disk at full resolution.

Tap any tile to play it on its own.

Sora 2 12

Twelve takes, 20 Aug. Frame 0 only — Sora pins the first frame and nothing else, so the ending had to be hoped for rather than composed.

SORA-01

Sora 2 · 720x1280 · 8s · 2026-08-20

conditioning: frame 0 only (input_reference)
from: S25 two-ingredients-pour

First pour attempt. Capsule reads as a still object with powder appearing near it.

SORA-02

Sora 2 · 720x1280 · 8s · 2026-08-20

conditioning: frame 0 only (input_reference)
from: S27 pour-physical

The raking-light regression. Backlight spilled onto the field and flattened the 11/255 gradient the whole section depends on.

SORA-03

Sora 2 · 720x1280 · 8s · 2026-08-20

conditioning: frame 0 only (input_reference)
from: S28 pour-final

Lighting split back apart. Motion better, but the halves still separate the wrong way round.

SORA-04

Sora 2 · 720x1280 · 8s · 2026-08-20

conditioning: frame 0 only (input_reference)
from: S29 opening-a

First take of the inverted capsule (body on top). Opening and pouring finally become the same motion.

SORA-05

Sora 2 · 720x1280 · 8s · 2026-08-20

conditioning: frame 0 only (input_reference)
from: S30 opening-b

Cap releases cleanly; powder still exits partly through the sealed end.

SORA-06

Sora 2 · 720x1280 · 8s · 2026-08-20

conditioning: frame 0 only (input_reference)
from: S31 opening-landed

Conditioned on a frame with the cap already down. Model had nothing left to author.

SORA-07

Sora 2 · 720x1280 · 8s · 2026-08-20

conditioning: frame 0 only (input_reference)
from: S32 opening-unbuilt

Mounds deliberately absent from frame 0 so they would have to build. Two takes run.

SORA-08

Sora 2 · 720x1280 · 8s · 2026-08-20

conditioning: frame 0 only (input_reference)
from: S32 opening-unbuilt

Mounds deliberately absent from frame 0 so they would have to build. Two takes run.

SORA-09

Sora 2 · 720x1280 · 8s · 2026-08-20

conditioning: frame 0 only (input_reference)
from: S33 opening-nascent

The take that went live on r12. Best of the Sora set: start-to-end change 8.51, mounds 2.52x.

SORA-10

Sora 2 · 720x1280 · 8s · 2026-08-20

conditioning: frame 0 only (input_reference)
from: S34 opening-dense

The dead take. Frame 0 was composed as a finished tableau, so almost nothing moves: 1.56 change, 1.00x mounds.

SORA-11

Sora 2 · 720x1280 · 12s · 2026-08-20

conditioning: frame 0 only (input_reference)
from: S34 opening-dense

The dead take. Frame 0 was composed as a finished tableau, so almost nothing moves: 1.56 change, 1.00x mounds.

SORA-12

Sora 2 · 720x1280 · 8s · 2026-08-20

conditioning: frame 0 only (input_reference)
from: S35 opening-inflight

Last Sora take. Its start frame is the one that eventually shipped, paired with a generated end frame.

Higgsfield · Seedance 6

Six takes, 20 Aug. First runs with a real start-and-end keyframe pair.

SEED-01

Higgsfield · Seedance · 1080x1920 · 2026-08-20

conditioning: start + end keyframes
from: hf start/end keyframe pair

First Higgsfield take. Whole run reframed away from the locked camera.

SEED-02

Higgsfield · Seedance · 1080x1920 · 2026-08-20

conditioning: start + end keyframes
from: hf start/end keyframe pair

12s, new end frame v3. Longer run, pour thins out in the back half.

SEED-03

Higgsfield · Seedance · 1080x1920 · 2026-08-20

conditioning: start + end keyframes
from: hf start/end keyframe pair

12s, same pair. Cap settles better; column still fans wide.

SEED-04

Higgsfield · Seedance · 1080x1920 · 2026-08-20

conditioning: start + end keyframes
from: hf start/end keyframe pair

8s against start v2. Tighter, but the capsule never fully empties.

SEED-05

Higgsfield · Seedance · 1080x1920 · 2026-08-20

conditioning: start + end keyframes
from: hf start/end keyframe pair

8s against start v3. Densest Seedance pour of the set.

SEED-06

Higgsfield · Seedance · 1080x1920 · 2026-08-20

conditioning: start + end keyframes
from: hf start/end keyframe pair

8s, end frame v5. Best Seedance cap behaviour; column still cones out.

Round-2 board 3

Three candidates, 21 Aug, built against a tighter brief: one column the whole way down, splitting into two touching mounds only at the floor.

PV-01

Higgsfield · MiniMax H3 · 1440x2560 · 6.6s · 2026-08-21

conditioning: start + end keyframes
from: round-2 keyframe pair

Rejected on the round-2 board: the column fans into a wide cone a third of the way down and the capsule never fully empties. Mound correlation +0.59.

PV-02 ROUND-2 PICK

Higgsfield · FLUX 3 · 1088x1920 · 6.0s · 2026-08-21

conditioning: start + end keyframes
from: round-2 keyframe pair

Round-2 winner on the tight-column brief. Holds one width the whole drop and parts only at the floor. Mound correlation +0.95 — the piles are visibly built BY the powder landing.

PV-03

Higgsfield · FLUX 3 · 1088x1920 · 8.0s · 2026-08-21

conditioning: start + end keyframes
from: round-2 keyframe pair

Same behaviour as PV-02, slower and calmer, thinner ribbon. Correlation +0.58. Would be sped up to ~6s rather than trimmed.

Shipped, superseded, and withdrawn 4

What made it onto the page — and the one that did not.

FINAL LIVE ON r13

Higgsfield · MiniMax H3 · 1440x2560 · 6.6s · 2026-08-22

conditioning: start + end keyframes
from: S35 opening-inflight + CHL-HOME-S37-opening-end

Shipped on r13. Best measured take: start-to-end change 14.25, mounds 4.26x, and it comes to rest unaided (tail motion 0.865 to 0.048) so it needed no retiming. 31 credits all in.

R13-WEB SERVED

FINAL, transcoded · 720x1280 · 6.6s · 605 KB · 2026-08-22

conditioning: n/a — transcode
from: what the homepage actually serves

The exact bytes section 02 plays today. Same motion as FINAL, downscaled and compressed for the page.

R12-WEB WAS LIVE

SORA-09 open-e, retimed · 720x1280 · 5.5s · 2026-08-21

conditioning: n/a — transcode
from: previous live clip

What was live on r12 before r13 replaced it. Sora open-e, trimmed and eased to a stop so it would not freeze mid-air.

REJECTED RULE BREAK

compose_pour.py (numpy) · 720x1280 · 5.7s · 2026-08-20

conditioning: VIOLATES the no-compositing rule
from: withdrawn

The composite you called a South Park skit — scaled mound sprites keyed over a generated plate. Rejected outright and the compositor was deleted from the project. Kept here only so the rule has a face: everything on screen is a full generation.