Both stickers are generated by the image model from your green-screen references and chroma-keyed, and both sit where the donor put its own sticker, measured in pixels rather than guessed.
| Donor | Sticker block | Top edge | Width | On screen | Running captions |
|---|---|---|---|---|---|
| 06 celinesfavoritefinds | x 107–971 of 1080, y 318 of 1920 | 0.166 | 0.800 | 0–2.8s | none at all |
| 02 natiscart | x 156–821 of 1080, y 400 of 1920 | 0.208 | 0.616 | 0–20.6s | 2 boxes against 364 other boxes |
Neither donor is a heavily edited video. Each is a selfie clip with one text sticker over the opening, and Script 6's donor carries no running captions whatsoever. Our ads now carry more post-production than the originals did, because the house rule is one caption look on every video.
Your six green-screen references are edit inputs for the image model, not pictures to copy by hand. The reference goes to google/nano-banana-2/edit with our words, comes back on the same flat green, and the green is keyed out on distance from (7,249,5) with a one-pixel erode. Two wrong turns before that: the retired swipes/callouts/ library, then the Python drawing script. Both are now blocked by a gate.
| Ad | Reference | Why that one |
|---|---|---|
| Script 6 | style2, two-tone | Two boxes for our two lines, and the boxes carry their own contrast over the light car headliner |
| Script 2 | style6, plain white | A two-line reference, and plain white reads on the dark car roof |
style5 was the first choice for Script 6 and failed twice: its reference is a five-line block, so the model kept all five lines and ghosted ours underneath. Match the reference's line count to the lines you need.