Most people treat the prompt box like a search bar. They type four words, hit enter, and then scroll through a wall of results that all look like they came from the same stock photo library. The model did what you asked — it just didn't have much to work with. The gap between a generic result and something you'd actually use usually comes down to whether you gave the model a scene or a label, and whether you set a single look instead of letting it average everything together.
Start with the subject doing something
"Woman in a field" gives the model almost no reason to choose one frame over another. "Woman in her thirties walking through tall dry grass, mid-stride, turning her head toward the camera, wind pushing her hair across her face" makes a dozen small decisions for it. Pose, direction of movement, and one gesture are cheap to write and they eliminate most of the boring outcomes before you ever see them.
A useful habit is to write the subject line as a half sentence of action and finish it with a single expression or gesture. If you can't describe what the subject is doing, the model will pick something neutral, and neutral is exactly what looks generic.
Set one light source and commit
Many weak results are actually two lighting setups fighting each other. You asked for a soft window feel but also mentioned "dramatic," and the model split the difference. Pick a direction (side, back, top), pick a quality (hard or soft), and say what the light is doing to the subject — "late afternoon sun from the left, hard shadow across the cheek" is a complete lighting brief. Once you have a setup that works, keep it identical across a series and only change what needs to change.
Trade adjectives for nouns
Words like "beautiful," "high quality," and "stunning" carry almost no information. Concrete nouns do the heavy lifting: material, era, lens feel, surface texture. "Linen shirt on a weathered oak table, 50mm feel, shallow focus" tells the model more than a paragraph of praise ever will. This habit alone tends to remove the plastic sheen that people complain about in their generated images.
Add one flaw on purpose
Clean is the default, and the default reads as fake. A crease in the fabric, a chipped corner on the mug, dust on the windowsill, uneven stitching — one or two small imperfections give the eye something to believe. Don't stack them. Pick the flaws that fit the story of the image and leave the rest tidy.
Change one thing at a time
The most common workflow mistake is rewriting the whole prompt after a miss. If the composition was right and the lighting was wrong, keep everything and edit only the lighting clause. Rewriting from scratch throws away the information the model already acted on, and you spend ten rounds rediscovering the same good pose.
Keep a short personal record of prompts that worked: the subject line, the light, the lens, the one texture or flaw. After a week or two you'll have a handful of building blocks you can recombine, which is faster and far more predictable than starting each session from an empty box. If you'd rather skip the blank page entirely, the generator on the homepage works fine as a place to test these blocks quickly before you get into a longer series.
Write the negative list last
Most models respond to what you don't want in some form, but the list only helps after the positive prompt is solid. Write the scene, the light, and the material first. Then add the few things that keep showing up uninvited — extra fingers, printed text, an over-saturated sky. Keep it short. A twelve-item avoidance list usually just confuses the model about what you actually asked for.
Give it a frame and a purpose
Telling the model what the image is for changes the output more than people expect. A portrait crop for a small circular avatar needs the face centered with breathing room on every side. A wide banner for a website header needs the subject off to one edge. Say the aspect and the crop intent out loud — "vertical, subject centered, space above the head" — and you'll stop throwing away good images because they were framed for the wrong slot.
None of this requires a longer prompt. Most of the time it's the same word count, rearranged so that every line answers a question the model would otherwise guess at. Start with one action, one light, one material, one flaw. That's usually enough to take an image from forgettable to usable.
When a result does land, save the exact prompt and the settings beside the image. Your future self will want to make twelve more of them, and a saved working prompt is worth more than any tutorial — including this one. Start with one clear subject and one committed light, and treat every prompt as a small set of decisions rather than a wish.