Seedream 4.5 prompt guide: patterns that hold up

Seedream 4.5 treats your input image as the baseline and your prompt as the delta. That single fact decides how you should write. Name what changes. Only describe what stays the same when it isn't obvious from context. The moment you re-describe the whole scene, you've invited the model to regenerate things you wanted untouched.

We tested this directly, edit versus re-describe, and the gap was not subtle: "keep the product, change only the background" preserved the source far better than a fresh description of the same scene ever did. The re-described version looked plausible on its own. Side by side with the source, the product had drifted.

When there's no input image, build the prompt in layers, in this order: subject and pose first, then style and color palette, then setting and lighting, then camera framing. The order matters because it mirrors what the model weighs most heavily. A camera note buried at the front costs you subject fidelity at the back.

Here's the same prompt built up layer by layer. Subject: "a woman in her 30s mid-laugh, head tilted back." Style: "shot like a premium lifestyle editorial, warm muted palette." Setting: "in a sunlit kitchen, late afternoon light through the window." Camera: "waist-up framing, shallow depth of field." Each layer narrows the space without fighting the previous one.

Keep one focal point. If you need legible text or a clean face in the shot, say so plainly and give it room. Competing focal points are where this model's famous text and face rendering starts to wobble.

Introduce each input's contribution explicitly: which image supplies the subject, which supplies a prop or style, and how they sit together in the new composition. The model responds to precise, layered instructions, and fusion is where that discipline pays most.

Fewer, cleaner inputs beat many marginal ones. Two or three sharp images that each clearly own one element fuse more reliably than a pile where several barely contribute. And keep lighting reasonably matched across inputs: one studio shot plus one dim phone photo forces the model to reconcile two qualities of material, which is where fusions start looking assembled instead of shot as one scene.

When one element of a fusion looks wrong and the rest is right, resist rewriting the whole prompt. Check that element's source image first, then sharpen only the instruction that refers to it.

Soft or under-detailed output on an edit usually traces back to the input image, not the wording. The model can sharpen and restyle, but it cannot invent detail that exists nowhere in the source. Before rewriting a prompt for the fourth time, look hard at what you fed it.

The same goes for faces in fusion work: if the source face is small, blurred, or half-turned, no phrasing rescues it. Swap the input for a cleaner one and the "prompt problem" tends to disappear.