Most people use image generators by typing something vague, hitting generate, and rolling the dice over and over. That works occasionally and wastes enormous time. This guide replaces the slot-machine method with a repeatable process. Follow these steps in order and you will produce intentional, controllable images instead of lucky accidents.
You do not need to understand all the underlying math to follow this. If you want the theory behind why each step works, the complete guide covers the mechanism. Here, we stay practical. Do this, then that.
Step 1: Define the Image Before You Type
Open a blank note, not the generator. Write down four things about the image you want: the subject, the style, the setting, and the mood. For example: subject is a woman drinking coffee, style is film photography, setting is a rainy cafe window, mood is calm and contemplative.
This thirty-second exercise prevents the most common failure: a vague prompt that the model fills with its own default choices. If you do not specify, the model decides, and it usually decides blandly.
Step 2: Write the Prompt in Layers
Assemble your prompt by stacking your four elements, most important first. The model weights earlier words slightly more heavily, so lead with the subject.
A reliable order:
- Subject and action first: "a woman sipping coffee, looking out a window"
- Setting next: "in a cozy cafe, rain streaking the glass"
- Style and medium: "35mm film photograph, shallow depth of field"
- Lighting and mood: "soft overcast light, calm and contemplative"
- Quality cues last, sparingly: "sharp focus, fine grain"
Avoid dumping fifty adjectives. Past a point, extra words dilute each other and the model loses the plot. Aim for a focused, well-ordered description.
Step 3: Add a Negative Prompt
Tell the model what to avoid. A standard negative prompt knocks out frequent defects: "blurry, distorted hands, extra fingers, watermark, text, lowres, oversaturated." Many tools support this directly. If yours does not, you can sometimes phrase exclusions in the main prompt, though dedicated negative prompts work better.
This single step removes a large share of the artifacts beginners spend hours fighting.
Step 4: Set Your Core Parameters
Before generating, configure the basics:
- Aspect ratio: match it to the intended use. Portrait for people, wide for landscapes, square for social posts.
- Steps: 25 to 40 is the sweet spot for most samplers. More rarely helps and costs time.
- Guidance scale (CFG): start around 7. Lower for more creative freedom, higher for stricter prompt adherence.
- Seed: leave it random for now. You will lock it later once you find a result worth refining.
Resist the urge to crank everything to maximum. High guidance and excessive steps often degrade results rather than improve them.
Step 5: Generate a Small Batch
Produce four images at once, not one. A batch lets you compare variations from the same prompt and spot which direction is closest to your vision. Looking at one image tells you little; looking at four reveals patterns in how the model is interpreting your words.
Scan the batch and pick the single image closest to what you want, even if it is imperfect. That becomes your anchor.
Step 6: Lock the Seed and Refine
Find the seed of your anchor image, usually displayed in the tool, and lock it. Now changes to your prompt produce controlled variations on that exact composition instead of an entirely new image every time.
With the seed fixed, make one change at a time:
- Adjust a single descriptor and regenerate
- Compare against the anchor
- Keep the change if it helps, revert if it hurts
This controlled, one-variable-at-a-time loop is how professionals dial in a specific result. Changing five things at once tells you nothing about what worked.
Step 7: Upscale and Fix
Once the composition is right, push resolution up. Most tools offer an upscaler that increases size while adding detail. Run it on your chosen image.
Then handle remaining defects:
- Inpainting lets you mask a flawed region, such as a malformed hand, and regenerate only that area
- Outpainting extends the image beyond its original borders if you need more room
- A final light pass in any editor can fix small blemishes the AI left behind
These touch-up tools separate a rough draft from a finished piece. For a tour of which tools handle each step best, see our best tools roundup.
Step 8: Save Your Recipe
When you land a result you love, save the full recipe: the exact prompt, negative prompt, seed, guidance scale, steps, and sampler. This is your reproducible formula. Next time you need a similar image, you start from a proven baseline instead of from scratch.
Building a personal library of working recipes compounds over time. Within a few months you will have go-to formulas for portraits, products, landscapes, and your common use cases. To avoid the errors that derail this process, keep our common mistakes guide handy.
A Worked Example From Start to Finish
To make the process concrete, here is a single image run all the way through. The goal: a portrait of a chef in a busy kitchen for a restaurant blog.
- Step 1, define: subject is a middle-aged chef plating a dish; style is editorial food photography; setting is a steamy professional kitchen; mood is focused and warm.
- Step 2, prompt: "a middle-aged chef carefully plating a dish, busy professional kitchen behind, editorial food photography, 50mm, shallow depth of field, warm tungsten light, focused expression."
- Step 3, negative: "blurry, distorted hands, extra fingers, watermark, text, lowres, oversaturated."
- Step 4, parameters: portrait ratio, 30 steps, CFG 7, random seed.
- Step 5, batch: four images; one has the right composition but a slightly muddy background.
- Step 6, refine: lock that seed, change only "busy professional kitchen behind" to "softly blurred stainless steel kitchen behind." Background cleans up; everything else holds.
- Step 7, fix and upscale: inpaint one hand that came out slightly off, then upscale to final resolution.
- Step 8, save: record the full recipe under "editorial portrait, warm kitchen."
Notice how few generations this took once the seed was locked. The discipline did the heavy lifting, not luck.
When to Break the Process
This is a default workflow, not a law. For throwaway, high-volume images where novelty beats precision, run a lighter version: a quick prompt, one batch, a fast pick, and skip the seed-locking refinement entirely. Matching effort to stakes is its own skill, and over-engineering a disposable social graphic wastes the time the process was meant to save. Save the full eight-step routine for images that genuinely matter.
Frequently Asked Questions
How many images should I generate before settling?
Generate in small batches of four and expect to run three to six batches with refinements before landing a keeper for anything important. Quick casual images may take one batch; polished work takes more iteration. The seed-locking step in this process cuts the total count dramatically.
What if locking the seed still gives me bad hands?
Use inpainting. Mask just the hands, write a short prompt for that region, and regenerate only that area, often several times. Fixing localized defects with inpainting is far more reliable than rerolling the whole image and hoping the hands come out right.
Should I write long prompts or short ones?
Medium length wins. Long enough to specify subject, setting, style, and mood, but not so long that descriptors compete and dilute. Aim for a focused paragraph, lead with the subject, and cut any word that does not change the image.
Do these steps work across different tools?
Yes. The underlying process, define, prompt in layers, negative prompt, set parameters, batch, lock seed, refine, upscale, is tool-agnostic. The specific buttons differ, but the workflow transfers across nearly every diffusion-based generator.
What is the single highest-impact step?
Step 6, locking the seed and changing one variable at a time. It converts random rerolling into deliberate control and is the habit that most separates frustrated beginners from confident users.
Key Takeaways
- Plan the image on paper before opening the generator
- Build prompts in layers, subject first, and keep them focused
- Always use a negative prompt to remove common defects
- Start with sensible parameters: ~30 steps, CFG around 7, random seed
- Generate batches, pick an anchor, then lock the seed and refine one variable at a time
- Upscale and inpaint to finish, then save the full recipe for reuse