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.
("I2I" is simply shorthand for image-to-image — you'll also see it written as image2image, img2img, or image 2 image.)
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.
T2I vs I2I at a glance
| Mode | Your input | The model treats it as | Best for |
|---|---|---|---|
| Text-to-image (T2I) | a written prompt only | the full brief | net-new visuals from nothing |
| Image-to-image (I2I) | one or more source images plus a prompt | structure and identity to preserve | editing, restyling, series consistency |
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.
Image-to-image on Molyin
Image-to-image is a standard mode across Molyin's image lineup rather than a single-model feature. What varies is how many references one generation accepts:
| Model | Reference images per generation |
|---|---|
| GPT Image 2 / GPT Image 2.5 | up to 16 |
| Nano Banana 2 | up to 14 |
| Nano Banana 2 Lite, Seedream 5.0 (Lite & Pro) | up to 10 |
| Nano Banana Pro, FLUX.2 Pro & Max | up to 8 |
| FLUX.2 Dev, Grok Imagine 2.0 | up to 5 |
Upload the source, describe the change in one specific instruction, and the generator shows the limit for the selected model plus any per-image surcharge before you generate.