AI Billboard Recreation (Grounding DINO + Stable Diffusion)

The problem

The client had a photo of a billboard and wanted an AI-generated recreation of the advert: detect the billboard in the image, extract its boundaries, upscale it to a higher resolution and change details in the artwork. A general image model cannot do that reliably from a street photo, because it does not know where the billboard ends and the sky begins, and left alone it rewrites the whole scene rather than the advert.

What I built

• Zero-shot billboard detection with Grounding DINO, which returns a bounding box from a text prompt without any training on billboard data.
• Extraction of the billboard region with OpenCV, using the detector's bounding-box coordinates to crop the advert out of the photo.
• A Stable Diffusion stage on the Realistic Vision V6.0 checkpoint that regenerates the cropped advert at 1280x1664, with ControlNet tile_resample holding the structure of the original and a denoising strength of 0.4 so only the details change.
• Ultimate SD Upscale with the 4x-UltraSharp upscaler working in 512-pixel tiles, plus soft inpainting on the masked area for seamless edges.
• A documented parameter set, including prompt, negative prompt, sampler, steps, CFG scale, seed and refiner, so the exact output can be reproduced.

Outcome

A pipeline that takes a street photo in and gives back a clean, higher-resolution recreation of the billboard with the requested changes, as shown in the before-and-after on the cover. Because every generation parameter is recorded, the client can rerun or adjust the result instead of depending on a lucky seed.