SpriteDX - Shot Detection 3

In previous posts, we looked into various methods to do shot detection. We then switched to collecting data for validating method.

We collected 256 samples of multi-shot animation and used TransNetV2 to detect shot boundaries automatically, then reviewed each sample to correct any mistakes.

After doing this, we realized that for 3-shot generations using Seedance, it clearly emerged that the shot boundaries are statically decided.

All of the correctly generated samples with 121 frames had:

• Shot 1: [0-40]

• Shot 2: [41-80]

• Shot 3: [81-120]

So, basically, for Seedance 1 Pro. It really doesn’t make any sense to “detect.“ We simply have to divide the frames evenly and we should have the best shot boundaries.

---

Without having collected data and doing some human labeling, I don’t think I would have realized this. So, perhaps it is more important to collect and study the data first rather than hypothesiszing.

— Sprited Dev 🌱