Editing and Variations

Editing

Part of: Multimodal AI: Vision & Images

Generating a brand-new image from text is easy. Changing one thing in a photo you already have, while leaving everything else untouched, is the hard and useful part. "Remove the trash can from this street photo but keep the building exactly as it is" is not a fresh generation problem; it is an editing problem. The trick is to give the model an existing image to start from instead of pure noise, and to tell it which pixels it is allowed to touch. What it is Three related operations all start from an input image rather than from scratch: - Image-to-image feeds an existing image plus a prompt; the model produces a new image that keeps the input's structure but applies your change. A strength dial controls how far it strays from the original. - Inpainting edits only a masked region. A mask marks which pixels may change; everything outside the mask is preserved exactly. - Variations generate new images that resemble an input without a text prompt, useful for "give me more like this." How it works Recall the denoising loop from the generation lesson. Editing changes where the loop starts . Instead of pure random noise, it begins from a noised version of your input image, so the structure

Challenge: The Inpaint Engine