01 / Start here
Begin with the image you want to see
Write the main subject and its action first. Then add only the setting, composition, light, and visual treatment that materially change the picture. A short natural-language sentence can be stronger than a long comma-separated list because it tells the model which details belong together.
For example, “a chair” leaves nearly every choice open. “An empty oak chair beside a north-facing studio window, three-quarter view, with soft daylight and a quiet editorial mood” gives the model a subject, spatial relationship, viewpoint, and light. It still leaves room for interpretation; a prompt is a direction, not a measured drawing.
One useful starting prompt
An empty oak chair beside a tall studio window, three-quarter view, soft daylight from the left, warm neutral room, with the chair clearly separated from the background.
Choose aspect ratio with the Generator’s shape control. Writing “16:9” in the prose is not a substitute for changing the image dimensions. You can adjust the prompt again before generating.
02 / Five visual decisions
Build the prompt from what changes the picture
You do not need every part in every prompt. Name the decisions that matter, and keep each one concrete enough that you could recognize it in the result.
- Subject
- Say who or what is present, what it is doing, and the feature that must make it recognizable. Put the main subject before secondary props.
- Setting
- Place the subject somewhere visible: a quiet studio, a wet city street, a kitchen counter, or an empty landscape. Include only details that should appear in frame.
- Composition
- Describe viewpoint, crop, placement, scale, and useful negative space. Use the real aspect-ratio control for canvas shape rather than relying on prose alone.
- Lighting
- Name a source and quality: soft window light from the left, a hard overhead sun, or a narrow rim light. Observable light is more useful than a vague word such as cinematic.
- Visual direction
- Choose medium, material, palette, texture, or mood: editorial photograph, ink illustration, muted stone and linen, or playful cut paper. Avoid stacking styles that fight each other.
03 / Learn by revision
Three examples, three different jobs
These examples show how a prompt changes when the task changes. They are editorial examples, not promises that a model will reproduce every detail exactly. The linked collections show actual visual directions and task-specific handoffs.
Create a new scene
Product photography
- Starting idea
- A ceramic mug for my shop.
- More useful direction
- A handmade blue ceramic mug on a pale stone tabletop, handle fully visible, soft window light from the left, clean background, with open space to the right for a product listing layout.
- What changed
- Material, placement, light, and empty space make the shot easier to evaluate. If the real mug has a particular glaze or logo, do not invent it in text; use a source image and review the output for fidelity.
Change an existing place
Interior design
- Starting idea
- Make this bedroom cozy.
- More useful direction
- Keep the room’s window placement and bed position. Change the palette to warm cream and muted timber, add a linen headboard and a low bedside lamp, and keep the morning light soft. Do not block the window.
- What changed
- This is a transformation instruction, not a text-to-image scene description. The source photograph carries the layout; the words separate what to keep from what to change. The result is a visual concept, not a renovation plan.
Make a focused edit
Photo editing
- Starting idea
- Fix the background.
- More useful direction
- Replace only the busy street behind the bicycle with a quiet pale wall. Keep the bicycle, its shadow, and the rider’s clothing unchanged; retain the original perspective and afternoon light.
- What changed
- A focused edit needs What, Where, and Keep. Naming the protected subject and lighting reduces ambiguity, but no prompt can guarantee pixel-perfect masking or identity preservation.
04 / Choose the right input
Use a source image when words should not invent the source
Text-to-image is useful when you are creating a new scene. If you already have a room, product, or photograph and the next result needs to relate to it, the source image is part of the instruction. Start an Image to Image transformation for a broader restyle, or use Photo Editor for a focused change.
Write the desired change and name what must remain recognizable. For a room: “Keep the ceiling, windows, and camera angle; change the wall finish and furniture mood.” For an edit: “Remove the sign behind the subject; keep the subject, shadows, and reflections.” Review the output against the source, especially faces, logos, text, and object edges.
If you want a new image inspired by a reference, Image to Prompt can draft words for visible qualities; it does not carry the original pixels. When the source must stay recognizable, upload it to Image to Image or Photo Editor instead. None of these tools promises exact geometry, readable text, or faithful identity, so compare every result with the original.
05 / Compare deliberately
Revise one visual decision at a time
Treat each result as feedback. The most efficient next prompt is usually a targeted correction, not a complete rewrite. Your generated result stays in Creation history, so you can compare instead of relying on memory.
- 01
Check the subject first
If the main subject is wrong or missing, move it to the first clause and remove competing subjects. Do not start by adding “ultra-detailed” or “masterpiece.”
- 02
Then check framing and negative space
Say “close crop,” “eye-level view,” or “subject on the left with empty space above” when placement matters. Use the actual aspect-ratio control for the canvas shape.
- 03
Change one light or mood decision
Replace “dramatic lighting” with a source and quality such as “soft side light from a window.” If a style suggestion pulls the result away from your goal, remove it.
- 04
Compare the result and keep the useful parts
A new generation may solve one issue while changing another. Compare the new image with the previous Creation and carry forward the details that worked. For reference edits, check whether protected areas stayed intact.
06 / Tool-specific, not magic words
Use model controls only in tools that expose them
Negative prompts, prompt weights, CFG, samplers, and LoRA are interface- and model-specific controls—not magic words to add here. This site’s generator does not expose them. The Stable Diffusion Reference explains what they do and where their behavior varies.
Open Stable Diffusion Reference →07 / Common questions
Before you press Generate
How long should an image prompt be?
Long enough to state the visual decisions that matter, short enough to avoid contradictions. Start with subject, setting, and one or two choices about composition or light; add detail only when a result shows why you need it.
Should I include “8K,” “masterpiece,” or an artist name?
Those words are not a general quality switch. Describe observable visual properties instead: medium, palette, material, crop, or lighting. Avoid relying on a named artist or website as a shortcut for a look you can describe directly.
Can a prompt preserve a real person, logo, or exact text?
No prompt alone guarantees those outcomes. Use an appropriate source-image workflow when you have a source, and review the output carefully. Do not publish an image that inaccurately represents a person or product.
What if I only have a rough idea?
The Image Generator includes a small Prompt Helper that suggests subject, setting, composition, light, mood, and aspect ratio. You can edit every suggestion and the final prompt before you generate.