Iterating without starting over

You have four images. One is nearly right. What happens next separates people who get what they wanted from people who generate ninety images and settle.

The seed is the starting noise

L1 said every image begins as random static. That randomness comes from a seed — a number. Same seed, same prompt, same settings, same image, every time.

That makes the seed the difference between steering and rerolling. Keep the seed and change one word, and you get the same image with that word changed. Change the seed and you get a fresh roll that happens to match your description.

Most tools expose this somewhere: a seed field, a “vary” button, an image-to-image mode, or a reference-image option. It is worth ten minutes finding where yours keeps it, because it converts the whole process from gambling into editing.

One variable at a time

The temptation with a nearly-right image is to fix everything at once: change the light, adjust the crop, add an object, switch the medium.

Then it improves, and you have no idea why — so you cannot do it again, and if it gets worse you cannot get back.

Change one thing. Keep a note of what you changed and whether it helped. Five deliberate iterations beat fifty rerolls, and at the end you have something more useful than an image: a list of what actually moves this model.

”Make it better” does nothing

An assistant can act on make this clearer because clarity has textual correlates. An image model cannot act on better, because it is not editing — it is generating from a description, and better describes no visual property.

Say the visible change:

…but shot from lower down. …but the light coming from behind the subject. …but in winter, bare branches.

If you cannot name what is wrong, look at the image and ask what you would tell a photographer to do differently. That sentence is your next prompt.

Know when to stop

Sometimes the composition is right and one detail is wrong — a hand, an object, a bit of text.

At that point stop generating. An image editor fixes a small defect in thirty seconds, where twenty more generations trying to prompt around it will change everything else you liked. Inpainting tools sit between the two: you mask the broken region and regenerate only that.

Rerolling a whole image to fix a corner of it is the most common way to lose a good result.

Try it now (6 minutes)

Generate an image you nearly like. Fix the seed if your tool allows it.

Now make three versions changing only the lighting. Then three changing only the camera position. Notice which axis moved the image more — that tells you where your attention belongs next time.

Check your understanding

1. Holding the seed constant while changing one word gives you: A completely different image The same image with that word's influence changed A higher-resolution version

2. Why change only one variable per iteration? Multiple changes take longer to generate Otherwise you cannot tell which change caused the improvement Models ignore all but the first change

3. When one small detail is wrong but the composition is right, you should: Reroll until that detail comes out correctly Fix it in an editor, or inpaint just that region Add 'perfect hands' to the prompt

Recap

Find where your tool keeps the seed, then change one variable at a time and note what moved. Name the visible change rather than asking for better, and when only a small region is wrong, edit or inpaint it instead of rerolling everything you already liked.