Writing prompts for image merging: specific beats poetic
Most bad results come from vague instructions, not weak models. Here's the four-part structure that fixes it, with before-and-after examples.
There's a persistent belief that prompting is a mystical skill. It isn't. It's technical writing, and it follows the same rule technical writing does: remove every sentence that could mean two things.
The four parts
Subject — who or what comes from which image. Placement — where things sit in relation to each other. Lighting — which source wins, and what the shadows do. Constraints — what must not change.
Miss any one of these and the model fills the gap with an average of everything it has ever seen. That average is rarely what you wanted.
Before and after
Before: "Combine these two photos of my dog and the park." After: "Place the dog from image one in the park from image two, standing on the grass in the middle distance, facing the camera. Keep his markings and collar exactly as they are. Match the overcast light from image two and put a soft shadow under his paws."
Before: "Make this product look good in a kitchen." After: "Place the bottle from image one on the wooden counter in image two, right of centre, upright. Preserve the label text and colours exactly. Add a tight contact shadow where the base meets the wood and a soft reflection on the counter surface. Morning window light from the left."
Neither rewrite is clever. Both are just unambiguous.
Words that pull their weight
Preserve, exactly, unchanged — these protect the things you care about. Match the light from image two — resolves the conflict before it happens. Contact shadow, cast shadow, reflection — name the physics you want. In the middle distance, left of frame, behind — placement, not vibes.
Words that waste space
Beautiful, stunning, amazing, high quality, 4k, masterpiece. They describe your feelings about the output, not the output. The model already knows you don't want an ugly picture.
When to stop rewriting
If two attempts with a specific prompt both come back wrong in the same way, the problem is upstream — usually a source image that's too small, too dark, or too heavily compressed. Fix the input rather than the sentence.