Edited by us
Image prompt — why the same person never comes back, and what seeds fix
Why character persistence, resolution, and style drift in image generation, and the procedure to write into the request.
By Mephify (operator)3 min read
What makes this hard
However precise the character description, the next image shows a different person. Across generations, identity is held by the seed, not the description. The description is merely necessary; without a seed procedure, faces change every cut even with an anchor.
Style is the same. Leave it unset and the engine renders differently on every run. And boilerplate like '8k, cinematic, best quality' is not direction. It layers a film look onto a request that wanted documentary realism, moving you further from the picture you meant.
What to add to the request
1. The seed is a procedure, not a value
Put it like this
Include this procedure in the prompt: generate 3–4 takes of the first cut, record the seed of the keeper, and reuse that exact seed for every later cut. If the tool takes reference images, also feed the first cut as reference.
A good seed cannot be known in advance, so it has to be a procedure rather than a recommended value.
2. Resolution comes from the use
Resolution is a value set by the use, not the engine's menu. Web posting: ~2048px on the long side; print: physical size at 300dpi; thumbnails: 1024px. State the use, take a recommended value, and add the line that the engine may not support it exactly.
3. Commit to one style and demand negatives
Specify style, lighting, composition, and color. Then separately demand a list of what must not appear (negatives). Likenesses of real people, and rights-encumbered elements like logos, characters, famous buildings, belong on that list.
The three parts of the finished prompt
- Main prompt — style, lighting, composition, color, and the character or brand anchor phrase
- Negative prompt — what must not appear. Likenesses of real people and rights-encumbered elements such as logos, characters, and famous buildings go here
- Parameters — aspect ratio, resolution, and the seed rule, written as plain values
When an anchor phrase is present, it is a series, and the seed procedure is never dropped. Someone making a single image has no reason to write an anchor. Waving it off with 'single generation, no seed needed' defeats the anchor's purpose.
What persists across a series, as values
With several images or variants, the elements that persist between them are declared as values. Persistent elements the anchor does not cover, such as time of day in the background, lens focal length and color temperature, are settled in the first image and cited by the later ones. Change only reads when something is known not to change.
Style also comes from two places: the art-style list and what you typed. When both exist, what you typed wins, and where they don't conflict they merge: the picked style sets light and lens, your words add texture. Boilerplate like '8k, cinematic, best quality' is never attached to any request. It isn't direction; it is a phrase that appears regardless of the request.
Checklist for writing it yourself
- 1If one person spans several images, did you require both the fixed description and the seed procedure
- 2Did you state the use and take a resolution recommendation
- 3Is exactly one style committed
- 4Is there a negatives list
- 5Are values written plainly, not in any tool's flag syntax? Write 'aspect ratio: 1:1', not '--ar 1:1'
The Image studio's 35 art-style chips are a quick-pick list of common styles. A style not on the list can just be typed. If a chip and your typed style conflict, what you typed wins.
Reading helps; trying is faster.