Gains:
- Ability to implement an iteration cycle consisting of discovery, selection, fine-tuning, local correction and formatting steps
- Ability to use seed fixing, negative prompt, inpainting and outpainting tools for the right problems
- Ability to produce consistent sets of images and avoid unnecessary iteration by setting a 'good enough' threshold
A single “perfect” image rarely comes on the first try. Professional design is finding an aspect and maturing it through iteration (repeated step-by-step improvement). In this unit, we will learn how to vary and fine-tune AI visuals in a controlled manner: we will move from "relying on luck" to "directing" with tools such as seed, negative prompt, inpainting.
Why variation and iteration?
Since the AI that produces images starts randomly, the same prompt gives different results each time. This is not a flaw, but an opportunity: you can choose the best among many options. But uncontrolled variation wastes time. The logic of iteration is this: produce a draft, determine what is good and what is bad, change the prompt or settings accordingly, produce again. With each round you bring the output closer to the target.
Key tools and terms
- Seed: A number that determines the random noise with which the image starts. Same seed + same prompt = same visual. If you fix the seed and change the prompt slightly, you will get "a variant of the same image with a slight difference". This is the golden tool for consistency.
- Negative prompt: The area where you write the things you do not want in the image ("bad hand, extra fingers, text, watermark, blur"). The fastest way to reduce defects.
- Inpainting (regional reproduction): Selecting only one region of the image (e.g. distorted hand) and having that region reproduced. You spot correction without reproducing the entire image.
- Outpainting: Enlarging the edges of the image outwards and creating new space; for example, expanding a square image into a horizontal banner.
- Image-to-image: Giving a starting image and saying "make it look like this but change that"; Ideal for changing the style while maintaining the composition.
- Denoise/strength: Determines how much the initial image will be preserved; Low value means little change, high value means lots of change.
Tip: When you capture a good image, note its seed. Then, by fixing that seed and experimenting with small changes (color, lighting), you can produce a consistent set for the brand.
Understanding the power of change (denoise)
The most confused setting when using image-to-image and inpainting is denoise/strength. It's a "how involved should I get involved?" Think of it like a dial. A lower value (e.g. 20-40%) largely preserves the initial image; It just makes small touches — perfect for shifting the color of a photo and adding subtle styling. A high value (e.g. 70-90%) produces a new image that is almost independent of the starting image; It can even change the composition. The wrong setting most often leads to two errors: setting the value too low and saying "nothing changed", or setting it too high and saying "the composition I liked disappeared". Practical method: start with a middle value (50%), adjust up or down depending on the result. Making three tries with the same prompt and seed by simply changing this dial will allow you to quickly find the right amount. This setting is where the difference between “trusting luck” and “leading” materializes: you decide how much to change, not the model.
Another powerful tool is image-to-image: giving an image you have (a sketch, a photo, a previous printout) as a starting point to the model and saying "keep its composition but change its style." Ideal for drawing a rough sketch and turning it into a finished illustration, or moving a reference layout to a different aesthetic. This tool moves AI from “creating from scratch” to “following your direction” and significantly increases control.
Step-by-step iteration loop
- Discovery tour: Write the prompt and produce 4 variations; Choose the most promising one.
- Pin the seed: Get the seed of the image you like.
- Improve point: Change the prompt with small touches (soften the light, shift the color) or add a negative prompt.
- Local fix: Repair the remaining defect (hand, edge, object) by inpainting.
- Format: Extend to different ratio by outpainting if necessary.
- Quality control: Enlarge, scan for defects, pass through brand filter.
three mini cases
Case 1 — Consistent set with Seed. A designer wanted 6 product images to be consistent for a cosmetics brand, but each production had different lighting. He fixed the seed of the image he liked and produced it by simply changing the product; The 6 images shared the same lighting and background. The ~2 hours he would spend on manual color correction was reduced to zero.
Case 2 — Recovery with Inpainting. In one illustration, the composition was perfect but the figure had a bad hand. Instead of reproducing the entire image, the designer selected only the hand area and corrected it with inpainting in 3 attempts. The problem was solved in 5 minutes without losing the popular composition.
Case 3 — Fault reduction by negative prompt. Unwanted text and watermarks were constantly appearing in an e-commerce image. The designer added "text, watermark, logo, signature" to the negative prompt; The rate of defective output dropped from 6 in 10 to 1 in 10. Sorting time was significantly shortened.
Case 4 — Adaptation with the power of change. A designer wanted to move a real product photo to a background compatible with the brand palette. With image-to-image, he first increased the power of change to 85%; The product has become unrecognizable. When I reduced the value to 35%, the product was preserved, only the background and lighting matched the brand. Finding the right dial in 3 tries saved hours of time compared to modeling the product from scratch.
Copiable templates
1) Discovery tour (multiple variations):
Create 4 variations for the following concept; have variety in composition and lighting, but maintain the brand tone (simple, warm). Concept: [prompt]. Negative: text, watermark, distorted hand, extra fingers, messy background.
2) Seed fixing and fine tuning:
Base this image (keep seed). Just make the following changes: soften the lighting a little more, make the background a tone lighter. Don't change the composition or product.
3) Inpainting instruction:
Reproduce the area (hand) I marked in the image: let it be a natural, five-fingered hand holding the product; keep the rest as it is.
4) Outpainting/ratio expansion:
Expand this 1:1 square image into a 16:9 horizontal banner. Produce matching background on the left and right; Keep the center element, leave plain space for text on the right side.
Weak prompt / Strong prompt
Weak: Fix this, he's bad-handed.
Result: The model reproduces the entire image, the liked composition disappears, new defects appear.
Powerful: Reproduce by inpainting only the hand area I marked: a natural five-fingered hand holding the cup; keep light and color in harmony with the environment; Don't touch the rest of the image.
Result: The liked image is preserved, only the incorrect area pinpointing is corrected.
Difference: Strong instruction separates the region, the desired result, and the area to be protected.
Tweaking tools table
vehicle
What does it do?
When to use
Seed fixing
Repeatability/consistency
Producing a compatible visual set
negative prompt
Exclude the undesirable
Text, watermark, anti-bad hand
inpainting
Regional fix
Repairing a single flaw
outpainting
Extend the edge
Changing ratio/format
Image-to-image
Keep the composition but change the style
Adapting an existing image
Denoise power
Change amount
Less/more intervention setting
Common mistakes
- Reproducing every flaw: Losing the popular image; However, spot correction is made with inpainting.
- Not taking note of the seed: Not being able to capture a good image again; Not being able to produce consistent sets.
- Leaving the negative prompt blank: Wasting time manually weeding out repetitive defects such as text, watermarks, faulty limbs, etc.
- Excessive iteration: Stuck on a good enough image and chasing small differences for hours; Set the “good enough” threshold.
- Force cropping the ratio afterwards: Losing the important element by trimming instead of outpainting.
In summary
Professional image production is not done in one shot, but in iterations: discover, select, fix the seed, fine-tune, localize, format, quality control. Provides seed consistency, negative prompt defect reduction, inpainting spot correction, outpainting format change. These tools move you from “relying on luck” to consciously moving the output closer to the goal. Set a "good enough" threshold and avoid unnecessary iteration.
Application task
Choose a concept, produce 4 variations and choose the best one. Note its seed and apply a small tweak (light or color). If there is a defect in the image (hand, edge, text), fix it with inpainting. Finally, outpaint the image to a different ratio (e.g. 1:1 to 9:16) and note which tool solves which problem throughout the process.
checklist
- [ ] I produced multiple variations on the discovery tour.
- [ ] I noted the seed of the image I liked.
- [ ] I excluded unwanted items with negative prompt.
- [ ] I fixed the defect by inpainting without reproducing it.
- [ ] I changed the ratio by outpainting when necessary.
- [ ] I stopped the iteration when I reached the "good enough" threshold.