Negative constraints

“No umbrellas” is a wish, and wishes are weak. A model renders what you describe — and naming a thing, even to forbid it, puts that thing in play. Negative constraints work when you convert them: describe the wanted state where the unwanted element cannot exist, or remove it afterwards with an edit.

Why wishes are weak

Every word you write is something the model can draw. “A beach with no elephants” mentions elephants; the words carry weight in the same space as everything else. Rendered results show it: forbidden objects drift back in, watermarks shade into “no watermark” prompts, and “don’t make it dark” produces very dark images. The wish states what you fear; the model renders what is stated.

Convert instead of forbid

Describe the state of the world where the unwanted thing does not exist:

• “no people” → “a deserted street at dawn”
• “no text” → “plain unmarked packaging, bare surfaces”
• “not blurry” → “sharp focus throughout”
• “no clutter” → “minimal composition with generous empty space”
• “don’t make it dark” → “bright, evenly lit”

Each conversion is renderable, checkable, and leaves nothing for the model to mis-draw.

What negative fields can and cannot do

Some tools — often those you run or fine-tune yourself — offer an explicit negative prompt field: a list of things to steer away from. Where offered, it behaves as a nudge rather than a law: keep it short and concrete (a handful of nouns, not paragraphs of wishes), and treat the output as probably clear, not provably clear. Many hosted tools have no such parameter at all; there, conversion is the whole toolkit. Either way, the reliable mechanism for true removal comes next.

Removal by editing

Where a tool supports edits, unwanted elements leave through the region: you supply the image, mark the area to change — a mask — and describe its replacement. Current editing APIs work this way: OpenAI’s image edits endpoint accepts “a mask that identifies the areas to replace”, and the tool’s own documentation notes that masking is entirely prompt-based — the model uses the mask as guidance but may not follow its exact shape with complete precision. So the professional loop for “make that thing gone” is: generate the scene close to right, then edit the region with a clear description of what should be there instead, then inspect that area closely. This is the same inpaint manoeuvre from the edit loop, aimed at subtraction instead of adjustment.

A bad example

Ask: “Office photo, no clutter, no people, don’t make it dark, no logos.”

Four wishes and one instruction about the thing being described. The result has a good chance of a person-shaped blur, a dark corner, and a logo-like smudge.

A better example

Ask: “A tidy sunlit office, empty desk with a closed laptop, bright even light, plain walls.”

Everything is a state the model can render — and the unwanted elements have nowhere to appear.

Why it works

Both halves are the track so far, pointed at removal. Conversion keeps you in the mode models handle best: rendering described states, not interpreting prohibitions. Editing handles what conversion misses: actual pixels get replaced in the actual region. Wishes do neither — they describe the absence of the preferred arrangement until it returns under another name.

Practice

Transfer

• Briefs in any tool: scan for “no / not / don’t” and convert each into the wanted state.
• Cleanliness in product shots: “bare, spotless, empty” beats “no mess” every time.
• True removal: mask the region and describe the replacement, then inspect closely.
• Anywhere you need certainty: the deliverable gets checked — wishes never shipped anything.

Next

That closes the image track — all twelve lessons. Put the whole set to work on the practice floor, or check where you stand with the skills check. The negative prompt entry sums up the term, and hallucination covers the plausible-looking detail this lesson keeps at bay.