Negative Prompts, Weights & Parameters

A negative prompt names what to push away from. Where the prompt pulls the image toward 'travertine, golden hour', the negative pushes it away from 'blurry, low quality, text, watermark, people, distorted windows, extra floors'. In Stable Diffusion it's a dedicated field; in Midjourney it's the --no flag ('--no people') or a negative-weighted term; in simpler UIs it's an 'exclude' box.

For architecture and interiors, a small standing negative prompt fixes a lot of recurring nonsense: warped mullions, melted railings, gibberish signage, phantom people, oversaturation. It's worth keeping a reusable negative block the way you keep a vocabulary - your default 'don't' list.

It helps to know why these artefacts recur. The model has seen countless photos where windows are slightly off, railings blur, and stray text sits in a corner, so those patterns lurk in its defaults and creep back unless you actively push against them. The negative prompt is you telling the denoiser which of its habits to resist. That's also why a good negative block is mostly the same across projects - the model's bad habits are consistent, so your corrective list can be too, trimmed slightly for each image.

One honest caveat: negatives nudge, they don't command. Pushing away from 'people' usually thins the crowd but may not empty the frame, and over-stuffing the negative can quietly drain life from the image (kill 'blur' too hard and you lose atmospheric depth). Use it as a corrective scalpel, not a wall.

Reordering (Lesson 1.1) re-weights terms for free, but when that's not enough you reach for explicit weights. The syntax differs by tool but the idea is identical: make one term louder or quieter than the rest.

In Stable Diffusion you wrap a term and multiply it: (travertine:1.4) turns it up, (people:0.5) turns it down - values above 1 amplify, below 1 attenuate. In Midjourney you use the double-colon: travertine::2 doubles its weight, and a negative number, people::-1, acts like a negative prompt. Firefly and other simple UIs lack a formal syntax, so you approximate by repeating and front-loading the term.

Weighting is how you resolve a tug-of-war. If 'brutalist concrete villa, lush garden' keeps burying the architecture under foliage, (brutalist concrete villa:1.3) or pushing 'garden' down restores the balance. The skill is restraint: crank a weight too high (past ~1.5 in SD, past ::3 in MJ) and the term takes over and the image degrades into caricature. Small moves, one at a time.

A useful discipline: try reordering before you reach for a weight. Because position already re-weights terms for free (Lesson 1.1), moving a term to the front of the prompt often does the job without any syntax at all, and it never risks the over-baked look a heavy explicit weight can cause. Treat explicit weights as the tool you reach for after reordering has failed - when two terms genuinely need to coexist but at different volumes. Beginners tend to skip straight to big weights and wonder why their images look forced; the fix is almost always a gentler hand and a better word order first.

Three more parameters shape the image without touching your words.

Aspect ratio frames the composition: --ar 3:2 for a classic architectural photo, --ar 16:9 for a cinematic establishing shot, --ar 2:3 for a portrait interior, --ar 1:1 for a study. It's not just cropping - the model composes differently for different frames, so a wide ratio invites context and a tall one invites a single vertical subject. Choose it to serve the shot, not as an afterthought.

Stylize (Midjourney --s, roughly a 'how much house-taste to apply' dial) trades fidelity to your prompt for the tool's own aesthetic polish. Low stylize hugs your words literally; high stylize makes prettier but looser images. Chaos (--c) sets how varied the initial grid of options is - low chaos gives four similar results, high chaos gives four wildly different ones, useful when you're still hunting a direction. Stable Diffusion's nearest equivalent is CFG scale (guidance): low CFG lets the model roam, high CFG clamps it tightly to the prompt (too high and it over-bakes). These are taste-and-variety dials - learn where your tool's defaults sit, then move deliberately.

The seed is the random number that sets the starting static (from Module 0.1). Two things make it the most important parameter in this lesson.

First, reproducibility: with the same prompt, same parameters and same seed, you get essentially the same image again. That turns generation from a one-off gamble into something you can return to and refine. Second, and this is the crux of the whole module: fixing the seed lets you isolate a single variable. Lock the seed, change one word or nudge one weight, and any difference in the output is caused by that change alone - because the underlying roll is held constant. Without a fixed seed you're changing the word and the roll at once, and you can't tell which caused what.

That's the bridge to Lesson 1.4's iteration method: seed-locking is the mechanism that makes disciplined, one-variable-at-a-time prompting possible. In Midjourney you read a job's seed and reuse it with --seed; in Stable Diffusion the seed is a first-class field you can copy and paste. Get in the habit of noting the seed of any image you like - it's the coordinate that lets you find your way back.

In practice you rarely touch one lever in isolation for final work; you set them together like a mixing desk. A realistic architectural steer might read: prompt text carrying your slotted words, (board-formed concrete:1.3) weighting the material you care about, a standing negative of people, text, watermark, distorted windows, --ar 3:2 for the frame, a moderate stylize so the model doesn't over-prettify past your brief, and --seed 42 locked so you can keep refining from a known roll. Each lever answers a different question - what, how hard, what to avoid, what shape, how much house-taste, which roll - and once you internalise the six questions the syntax of any specific tool is just translation.

The order of operations matters too. Get the words right first (Lessons 1.1 and 1.2); reordering is free and fixes more than beginners expect. Only then reach for weights, and only when a genuine tug-of-war needs rebalancing. Add the negative prompt to clean recurring artefacts, set the aspect ratio to serve the shot, and lock the seed the moment you have something worth refining. Reaching for heavy weights before you've fixed the wording is a classic beginner move that just buries a weak prompt under force.

But here is the boundary that defines the whole module. Every lever in this lesson is a nudge on a probabilistic process - they refine and bias, they never command. You cannot weight your way to a specific floor plan, negative-prompt your way to a precise structural bay, or seed your way to an exact existing building. When you need the model to obey geometry rather than lean toward an appearance, parameters have run out and you need conditioning - ControlNet, img2img, sketch input. That is Module 3, and recognising this boundary is itself a professional skill: it tells you when to stop fiddling with dials and change tools.