Image-to-image (I2I) is a generative AI technique that transforms an existing image into a new one, guided by a text prompt and optionally more reference images. Instead of starting from a blank canvas, the model starts from your picture — keeping what you want kept and changing what you ask it to change.
How it works
The model conditions its diffusion process on your input image as well as your prompt: the source provides structure, identity, and layout, while the prompt steers what should change. Modern instruction-tuned editors take this further — they accept plain-language edit commands ("replace the background with a rainy street, keep the subject unchanged") and multiple reference images at once, so a face, product, or art style stays consistent across every output.
What it's best at
Image-to-image is the workhorse of AI-assisted editing:
- Editing without masks — swap backgrounds, change outfits, remove objects, or relight a scene with a sentence instead of a selection tool.
- Style transfer — rerender a photo as watercolor, anime, oil paint, or blueprint while preserving the composition.
- Consistency across a series — feed the same reference images to keep a character or product identical across campaign shots.
- Upgrading rough inputs — turn a sketch, screenshot, or low-effort render into a polished visual.
What a good input looks like
Start with a clean, well-lit source image, then write the prompt in two halves: what must stay (the subject, the pose, the brand colors) and what should change (the background, the style, the season). The more reference images you provide, the more faithfully identity survives the transformation.