When the image comes out wrong

Something came back wrong. There's a ring around the artwork nobody asked for, or the creature looks bored, or the word is spelled with two Ts.

The reflex is the same every time: add a line to the prompt telling it not to. no border, no circle, no seal, not blurry, no extra limbs. The list grows every session and never gets shorter, because nothing ever proves a line on it was useless.

That reflex is the most expensive habit in prompting, and it's usually treating the wrong layer.

Why doesn't adding more negatives help?

Because attention is finite, and every prohibition spends some of it.

This is documented outside as well as inside: past a certain length, negative lists show clearly diminishing returns, start conflicting with each other, and can suppress detail you actually wanted. The current consensus sits around a handful of carefully chosen terms rather than an inherited wall of them — and the effect is weaker on newer architectures than it was on the models the habit was formed on.

The rule that survives contact with production is narrower than the habit: only treat a disease you have actually seen. A negative earns its place by having fixed an observed failure. Adding one preventively, for a problem this image never had, dilutes the ones doing real work.

So what is actually broken?

Four different layers fail in ways that look similar and want completely different fixes.

• Structure — the composition is wrong: wrong crop, wrong arrangement, elements in the wrong relationship. Fix the geometry.
• Slot — the right composition containing the wrong object. Fix the noun.
• Texture — the whole thing has drifted out of the style you wanted. Fix the anchor.
• Chance — spelling, a stray limb, a bad roll. Fix nothing.

Naming the layer before editing is most of the skill. The ring around the artwork is a structure problem; a bored creature is a slot problem. Neither is improved by a prohibition, and adding one makes the next diagnosis harder because now you can't tell which line is load-bearing.

Why don't adjectives make it fiercer?

Because they mostly don't land, and that's measurable rather than aesthetic.

In our own production testing, concrete noun props were delivered close to one-for-one, while adjectives like hellish or aggressive did essentially nothing. Asking for a fiercer beast returns the same beast. Asking for strings of saliva, a snapped chain on a spiked collar, smoke coming off the fur, and no pupils returns a different animal entirely.

So intensity is a list, not a modifier. When something reads as flat, the move is to add objects — things that could be photographed — rather than to escalate the description of a mood. The same test showed the parallel case for composition: verifiable geometric language is obeyed, and feeling words are not. Arched upward in a strong curve works. Dynamic layout doesn't.

What if it renders the wrong object entirely?

Swap the noun rather than banning the thing you got.

A measured example from our own batches: asking for crossed dumbbells inside a woodcut style reliably produced wooden mallets. The instinct is to add no mallets. The fix that worked was naming the object more precisely — cast-iron dumbbells with round weight plates — because the model wasn't disobeying, it was resolving an under-specified word inside the style it had been anchored to.

That's the general shape. A wrong object is nearly always an under-described object, and a prohibition leaves the vagueness in place while adding noise on top of it. Describing the thing better fixes the cause; banning what arrived treats the symptom.

When should I change nothing at all?

When the failure was stochastic. Spelling errors, an extra finger, one bad composition in an otherwise good set — those are draws, not instructions.

Editing the prompt in response to a random failure is how a good prompt slowly turns into a bad one. You add a clause, the next roll happens to be fine, and now that clause is permanently in your prompt with a false claim to credit. Repeat that for a month and you have a long, brittle prompt where most lines are superstition.

The correct response to a bad roll is another roll.

Why change only one thing at a time?

Because the layers have wildly different leverage, and changing several at once destroys the evidence.

Structural changes move the result most; swapping props is smaller but lands almost every time; changing the style anchor shifts everything globally but indirectly, mostly affecting how the lettering and overall texture read. If you turn all three knobs and the result improves, you've learned nothing you can reuse.

One knob per round is slower per attempt and much faster overall, for the same reason a graded set beats re-rolling — both are ways of keeping the evidence intact while you move.

Do I have to manage any of this?

Not for the print-asset side of it, which is where most inherited ban lists come from.

Generating here applies its own fixed requirements to every image — a cut-out-ready background, the subject centred with margin, no garment or mockup in frame, legible at thumbnail size. That's the part of the ban list that exists to serve the pipeline rather than the picture, and it's already handled, permanently, rather than being retyped and half-remembered each session.

What's left is the part that was always the interesting half: which objects are in the frame, and how they're arranged.

How do I get one made?

Describe the idea at JustOG and the directions come back as a set. Pick one, drag the crop frame, see it composited on the real garment, and it's made to order and shipped.

Designs other people have published are in the shop.

Before adding another no, name the layer. It's usually a noun.