If you use image-to-image tools for product shots or concept art, one habit makes results far more predictable: split your instruction into a KEEP list and a CHANGE list. The Seedream 5.5 image to image page basically recommends this ("explain what to keep and what to change in the new image"), so here is how it looks in practice.
Summary of the linked page: seedream5-5[.]art/image-to-image (read 2026-10-31)
How the workflow is described
Upload a reference or pick one from your library, say what to retain and what to change, set the composition and canvas ratio, then review the required credits and submit. Afterwards you can open the result, download it, favourite it, or use it again as a reference with a revised prompt. References can be JPG or PNG files or images from your library.
Why two lists help
A single sentence like "make this sneaker photo look better for autumn" leaves the model to guess what "better" means, and it may change the shoe itself. Two lists remove the guessing.
Self-made diagram: a vague instruction versus keep / change lists
KEEP should name the things a client would notice if they moved: product shape, colourway, logo position, camera angle. CHANGE should be concrete and visible: "fallen leaves on wet stone", "warm low light", not "autumn vibe". The page's own FAQ says a focused prompt makes the role of each reference easier to understand, which is the same idea.
Pick the canvas for the destination
The FAQ gives simple ratio guidance, which is worth deciding before you write the prompt, because composition depends on it.
Self-made diagram: canvas choice by where the image will be used
Iterate one idea at a time
The page's how-to opens with "refine the visual one idea at a time". After a good result, use it as the next reference and change only one thing: the light, or the background, or the viewpoint. If the next result is worse, you know exactly which change caused it.
Self-made diagram: checks before reusing a result
Template
KEEP: (shape, colours, logo/text, angle)
CHANGE: (one or two concrete, visible changes)
CANVAS: (square / portrait / landscape, based on destination)
Interested to hear if anyone here uses a different structure for edits, especially for character art where "keep the face" is harder to define.
Summary of the linked page: seedream5-5[.]art/image-to-image (read 2026-10-31)
How the workflow is described
Upload a reference or pick one from your library, say what to retain and what to change, set the composition and canvas ratio, then review the required credits and submit. Afterwards you can open the result, download it, favourite it, or use it again as a reference with a revised prompt. References can be JPG or PNG files or images from your library.
Why two lists help
A single sentence like "make this sneaker photo look better for autumn" leaves the model to guess what "better" means, and it may change the shoe itself. Two lists remove the guessing.
Self-made diagram: a vague instruction versus keep / change lists
KEEP should name the things a client would notice if they moved: product shape, colourway, logo position, camera angle. CHANGE should be concrete and visible: "fallen leaves on wet stone", "warm low light", not "autumn vibe". The page's own FAQ says a focused prompt makes the role of each reference easier to understand, which is the same idea.
Pick the canvas for the destination
The FAQ gives simple ratio guidance, which is worth deciding before you write the prompt, because composition depends on it.
Self-made diagram: canvas choice by where the image will be used
Iterate one idea at a time
The page's how-to opens with "refine the visual one idea at a time". After a good result, use it as the next reference and change only one thing: the light, or the background, or the viewpoint. If the next result is worse, you know exactly which change caused it.
Self-made diagram: checks before reusing a result
Template
KEEP: (shape, colours, logo/text, angle)
CHANGE: (one or two concrete, visible changes)
CANVAS: (square / portrait / landscape, based on destination)
Interested to hear if anyone here uses a different structure for edits, especially for character art where "keep the face" is harder to define.