Some templates earn their keep in the pre-processing. Our latest one — a "Santa caught on camera" clip — started as a prompt experiment: ask a video model for 1998-CCTV-style footage of Santa mid-delivery in your living room, and it politely ignores the style instructions. Art direction does not survive prompting; it has to be baked into the input image.
The grade happens in the browser
The moment a photo is picked, we run it through a canvas pipeline before anything uploads:
draw the photo and desaturate it to luminance-only,
push the contrast up,
lay down per-pixel noise for the grain,
burn a soft vignette into the corners.
The graded JPEG — not the original — is what the video model receives. A side effect we like: the thumbnail in the picker is pixel-identical to the model's input, so there is no "why does my clip look like that" mystery later.
Two prompt lessons we kept
You cannot prompt your way to a style; you can prompt your way to a line of dialogue.…