One Image, Twelve Attempts: What Each Edit Actually Did

Normally the trail evaporates. You try things, one of them works, and by the evening you could not reconstruct how you got there. So on one image I wrote everything down: the prompt at each step, what I changed, and what came back. Twelve attempts, roughly eighty cents, and the useful part is that only seven of them were worth anything.

The target: a man in a workshop, evening, something that felt like a real place rather than a stock photo of Craftsmanship.

Attempts 1 to 3: writing wishes

1.“A craftsman working in his workshop, warm atmosphere, cinematic, highly detailed”

Came back as exactly what you are imagining: a bearded man in an apron, evenly lit, in a workshop with no dust in it. Technically fine, completely generic. The problem is that four of those seven words are mood, not content.

2.Added “golden hour, moody, professional photography”. Got the same picture, slightly more orange. This is the attempt where I should have noticed I was adding rather than replacing, and did not.

3.Added “8k, masterpiece”. Nothing changed. Nothing was ever going to change. Two attempts of pure waste, and I include them because they are the most common two attempts anybody makes.

Attempt 4: the first real edit

4.“A man in his sixties repairing a chair in a woodworking shop, low evening light through a single window on the left, long shadows across the floor, sawdust in the air, waist-up, 50mm, photographic”

This is a different picture. Not because it is longer, but because every clause is now something a person could act on: an age, an action, one named light source with a direction, a specific piece of atmosphere, a framing, a lens.

The sawdust is the clause I would keep if I could only keep one. It does more for believability than any adjective, because it implies the room is being used rather than photographed.

Attempts 5 to 8: one variable at a time

Here is where it became useful, because from this point I changed exactly one thing per attempt and kept the rest byte for byte.

5.Window on the left became “window behind him”. Big change, worse: he became a silhouette and the face was gone. Useful failure, because it showed me how much the light clause was carrying.

6.Back to the left window, added “a work lamp on the bench, warmer than the window”. Two sources that disagree. This is the attempt that made the picture look like a place. Everything after it was refinement.

7.Changed “waist-up” to “wide, showing the whole bench”. Worse for this idea. The room got interesting and the man stopped mattering. Reverted.

8.Added “visible skin texture, hands with real wear”. Small change, disproportionate effect. Hands are usually where a render gives itself away, and asking for the wear rather than hoping for it helped.

Attempts 9 to 12: the part I should have skipped

By 8 I had the picture. What I did next was fiddle.

9 to 12were: a different shirt colour, then a different chair, then “slightly lower angle”, then back to the version from 8 because it had been better all along. Four attempts, about twenty-six cents, zero information gained.

The tell was there and I ignored it. Attempts 9 and 10 came back different but not better, and that is the signal. Different-but-not-better means you have stopped steering and started sampling, and every further attempt is a coin toss you are paying for. When two consecutive edits fail that test, the image is done and you are just reluctant to admit it.

What the trail actually taught

• Attempts 1 to 3 changed nothing because they added adjectives. Adding is not editing.
• Attempt 4 changed everything because it replaced mood words with things a person could act on.
• Attempts 5 and 7 were both failures and both worth doing, because a reverted change tells you what a clause is carrying.
• Attempt 6 was the one that mattered, and it was a second light source. It usually is.
• Attempts 9 to 12 were fear of stopping.

One more thing worth saying because it goes against the grain of iterating: sometimes the right move is to throw the prompt away. If the composition is wrong rather than the details, twelve edits to a bad premise will not save it, and a fresh sentence written after you have decided what you actually want beats all of them. Iterating is for when the shape is right.

If you want the assembled version of what attempt 4 discovered, five complete prompts are laid out with the clause order explained, and the post on why images look generated covers why the second light source does so much. Both of them are the tidy version of this mess. The mess is how it actually goes.

Save your prompts, not your images. The text is where the lesson lives, and files only stick around for thirty days anyway. Then go and waste some attempts deliberately, one clause at a time.