Editing, Upscaling and Consistency

Editing techniques

• Inpainting: select an area and regenerate only that part (fix a hand, change a shirt colour, remove an object).
• Outpainting / generative expand: extend the canvas beyond the original edges, for example turning a square image into a wide banner.
• Background removal and replacement: isolate a product and place it in new scenes.
• Image-to-image: start from a sketch or photo and restyle it, with a strength setting controlling how much changes.

Upscaling

Generated images are often limited in resolution. AI upscalers increase resolution while adding plausible detail. For print or large displays, upscale, then inspect fine details carefully; upscaling can invent texture.

Real products: photograph, then composite

For e-commerce, customers must see what they will receive. Photograph the real product, remove the background, and use AI for scenes and lighting around it. Do not generate the product itself if details (logos, stitching, colours) must be exact.

Consistent characters and styles

• Use the tool's character or style reference features, or a fixed seed with small prompt changes.
• With open models, train a small LoRA on 15–30 images of a character or product (with rights to those images) for strong consistency.
• Write a style guide prompt fragment (palette, lighting, medium) and reuse it for every image in a campaign.

Text in images

Some models render short text well; many do not. For logos and important text, add them afterwards in a design tool (Figma, Canva, Photoshop) for accuracy.

Exercise

Take a photo of a real object, remove the background, generate three lifestyle scenes around it with consistent lighting, outpaint one into a 16:9 banner and add a headline in a design tool.

Check yourself

• What is the difference between inpainting and outpainting?
• Why photograph real products instead of generating them?
• How can you keep a character consistent across images?