Almost everyone hits this wall in the same way. You write a careful prompt, fifteen lovely descriptive words, and the model proudly delivers nine of them while quietly ignoring the rest. You wanted emerald green eyes and you got hazel. You asked for soft window light and got a studio flash. You added "red dress" and now the entire image, walls, floor, mood, has gone red. The instinct is to repeat yourself louder, to cram the word in three more times, but that almost never works the way you hope. The fix is not more words. It is telling the model how much each word is worth.
That is the whole idea behind prompt weighting. Under the hood, a diffusion model reads your prompt as a field of attention, and some tokens naturally grab more of it than others. Color words are bullies. Strong nouns shout over soft adjectives. Anything near the front of the prompt tends to matter more than anything near the end. Weighting is the syntax that lets you override those defaults on purpose, pushing a quiet detail forward and pulling a loud one back, until the spread of emphasis matches the picture in your head.
The Syntax: Numbers Beat Repetition
Most tools that support weighting use a parentheses-and-number convention. You wrap a term in parentheses and assign it a multiplier, where 1.0 is neutral, anything above 1.0 boosts the term, and anything below 1.0 dials it down. In the most common syntax it looks like (emerald green eyes:1.4) to emphasize, or (background:0.6) to suppress. The number is the volume knob for that idea.
| You write | What it does |
|---|---|
| (term:1.3) | Boosts the term moderately, a reliable first nudge |
| (term:1.5) | Strong emphasis, use when a detail keeps getting ignored |
| (term:0.7) | Quiets a term that is overpowering the rest of the image |
| (term:0.4) | Heavy suppression, almost a polite way of saying "barely" |
Note the range that actually works. The sweet spot for a boost lives between roughly 1.1 and 1.5. Crank a term to 1.8 or 2.0 and you do not get "even more of it," you get distortion, blown-out color, melted anatomy, the model overcommitting to one idea at the expense of coherence. Weighting rewards a gentle hand. Think of it as seasoning, not flooding.
The mistake nobody warns you about: typing a word three times is not 3x weight, it is three chances to confuse the tokenizer. One clean term at (word:1.4) beats "word, word, word" every single time.
Boosting Versus Dialing Down
Here is the shift in thinking that makes weighting click. Beginners only ever reach for the boost, trying to force a stubborn detail to appear. But half the real power is in suppression. When one element is eating your whole image, the elegant move is not to fight it with five competing boosts, it is to turn that one element down.
Boost when a detail keeps vanishing
If you keep asking for freckles and the model keeps giving you flawless skin, the term is losing the attention fight. A small boost, freckles at 1.3, pushes it back into contention without warping the face. The same trick rescues any subtle thing the model treats as optional: a specific fabric texture, a piece of jewelry, the direction of a gaze.
Suppress when one term takes over
Color bleed is the classic example. You write "red dress" and the room turns into a darkroom. Dropping it to (red dress:0.85) often keeps the dress red while letting the rest of the scene breathe. Suppression is also how you tame a setting that keeps dominating the subject, or a stylistic word like "dramatic" that the model loves a little too much.
A Few Habits That Keep It From Backfiring
- Change one weight at a time. If you adjust three terms at once and the image improves, you have learned nothing about which change did it. Move one slider, keep the seed locked, and watch what shifts. This pairs perfectly with reusing a fixed seed, which we walked through in our seeds guide.
- Start small and climb. Reach for 1.2 before 1.5. You can always add more emphasis, but you cannot un-melt a face after the fact. Most "fixed" prompts only needed a 0.2 nudge, not a dramatic one.
- Read the whole prompt as a balance. Every boost you add quietly steals attention from everything you did not weight. If you push five terms up, you have effectively pushed everything else down. Weight the two or three things that truly matter and leave the rest neutral.
- Mind the front of your prompt. Word order already acts like soft weighting, so the things you care about most can simply live earlier in the sentence. Sometimes the cleanest fix is moving a term forward rather than wrapping it in a number at all.
Try This Tonight
Take a prompt that has been almost working, the one with the detail you can never quite land. Lock your seed so nothing else moves. Find the single term the model keeps ignoring and wrap it at 1.3, then regenerate. If it shows up but now overpowers, drop it to 1.15. If something else has started bullying the frame, knock that down to 0.8. You are not rewriting the prompt, you are mixing it, sliding two or three faders until the balance matches your intent. That is the entire skill, and once it clicks you will stop seeing prompts as lists of words and start seeing them as a budget you get to spend on purpose.