Ce guide est en anglais
How to find a seed that ends well
Mis à jour le 13 septembre 2026
The bottleneck for anyone making physics videos is not generating a video — it is generating a good one. Most end anticlimactically: the ball escapes too early, or nothing happens for thirty seconds. Simbounce attacks exactly that: because simulating without drawing is orders of magnitude faster than rendering, the editor tests dozens of variations in the background and shows which ones end well.
Pas à pas
- 01
Set the scene and parameters the way you want them
The search evaluates the recipe as it is. Change the parameters before, not after: a seed that ends well with a 20-degree gap has no reason to end well with 40.
- 02
Open the Good seeds panel and start the search
The panel says what it does: it simulates without drawing and ranks by outcome, and it never spends credit. The search runs in the background; the preview keeps responding while it runs.
- 03
Read the candidates
Each candidate shows the number of collisions, which step holds the climax, and a score from 0 to 100. The climax is the moment of highest event density — and a video that retains has that peak near the end, not in the middle.
- 04
Choose by the shape of the ending, not the score alone
Between two similar candidates, prefer one that escapes or settles near the end over one that times out. Timing out means the scene did not finish within the maximum duration: the video ends on the clock, not on the story.
- 05
Apply the seed and check the preview
Clicking a candidate loads the seed into the editor. Watch it once all the way through before exporting: the search measures events, not beauty, and the eye still decides.
What the score measures
Four things, weighted: how close to the end the climax lands, how many collisions happen relative to a target, how regular the rhythm between them is, and how close the duration is to a fifteen-second target. Videos that only time out are rejected by default, because they are rarely interesting.
The scene pages on this site show, for every scene, the five best seeds of a fixed sample evaluated with the same heuristic — and the share of variations that escape, settle or time out with the default parameters. It is a good starting point for knowing what to expect from each scene.
When no seed passes the threshold
The search has a threshold, and sometimes no candidate crosses it. That is information about the parameters, not about the search: a gap that is too small in an escape scene only produces timeouts; gravity that is too high settles everything early. Change one parameter and search again.
Some scenes have no natural ending by design — the pendulum wave and the Fourier drawing are made to loop. In those, timing out is the expected behavior, and what decides the duration is what you choose in the export.
Why it costs nothing
Evaluating a seed means running the physics without drawing a single pixel, in your browser. There is no server, no queue and no credit involved. Actually rendering — in the browser or on the server — is the expensive step, and that is why the search exists: so you only render what you already know works.