Ever ask @ChatGPT to fix one tiny detail in an image and get back a whole new picture? OpenAI's GPT Image 2.6 guide has some practical tips for getting the image you want and keeping the parts you already like. 1. Say what you're making. A wide launch banner, a portrait, and a product photo need different layouts. Give the model the format and say where the main subject should go. 2. Describe what you'd actually see. "A rocket on the right, pink light on the exhaust, and dark sky on the left for a headline" gives the model more direction than "epic synthwave rocket." 3. Put exact text in quotes. Try "Write 'Omarchy 4.5' once at the top in large white letters." Then zoom in and check the spelling and legibility before posting. 4. Edit one thing at a time. Use the last good image and name what should stay fixed. "Remove the letter from the rocket. Keep the logo above it, the colors, and the framing." If a logo needs to be pixel-perfect, paste the original back in after the edit. 5. Give every reference image a job. Tell the model to use image 1 for the layout, image 2 for the logo, and image 3 for the color palette. That makes it clear what to take from each one. 6. For API edits, circle the target. GPT Image 2.6 accepts a transparent markup layer, so you can point to the exact object you want changed and ask for the circle to disappear in the final image. 7. Tune planning and detail separately in the API. Try low complexity for a simple edit. If the layout looks right but the small print is blurry, raise quality. Start with Flare when speed matters, and try Sunburst for demanding layouts or precise edits.