---
title: "Image to image"
description: "Using an existing picture as the starting point for generation — how much of the original survives, and when this is the only mode that will work."
url: "https://higgsfield.wiki/image-to-image/"
verified: "2026-08-26"
publisher: "Higgsfield Wiki — independent reference, not affiliated with Higgsfield AI"
---

# Image to image

Using an existing picture as the starting point for generation — how much of the original survives, and when this is the only mode that will work.

Image-to-image starts from a picture you supply rather than from noise. The output is anchored to that input, which is what makes it the controllable mode.

## Why it matters more than text-to-image

Text-to-image is a slot machine with a very large prize pool. Image-to-image is a tool. If anything specific has to appear in the output — a real product, a particular face, a layout you already designed — supplying it beats describing it every time, because description cannot carry identity.

## How much of the original survives

Most implementations expose a strength or denoise control, and understanding its shape saves a lot of trial and error:

| Strength | What happens | Use for |  |

| Low | Output closely resembles the input; small changes only | Colour grading, light retouching, subtle style shifts |  |

| Medium | Composition and subject survive; surface and detail change | Restyling, changing materials, era or mood shifts |  |

| High | Only rough composition survives; content is reinvented | Using an image as a loose layout sketch |  |

The most common mistake is running high strength and then complaining the subject changed. If identity must survive, you want low to medium, plus a model that handles identity well.

## Practical uses

  - **Restyling a photo** — same scene, different rendering.
  - **Fixing one region** — where inpainting is available, regenerate a selection and leave the rest untouched.
  - **Producing a first frame for video.** This is the highest-value use. Get a still exactly right, cheaply, then animate it — see [image to video](/image-to-video/).
  - **Iterating a design** — feed the output back in and nudge repeatedly, rather than re-rolling from scratch.

## Input quality decides output quality

The model amplifies what you give it. A soft, low-resolution, already-retouched input yields a soft, plastic output no matter how good the prompt is. Start from the sharpest, most neutral version you have — an unedited original beats a filtered export.

For instruction-style editing where you name a change in words and everything else stays put, see [Nano Banana](/models/nano-banana/), which is built for that.

## Common questions

### What is the difference from inpainting?

Image-to-image transforms the whole frame. Inpainting regenerates only a region you mask and leaves the rest of the pixels alone.

### Why did my subject change?

Strength was too high. Lower it so more of the original survives, or use a model built for identity preservation.

### Can I use it to change only the background?

That is better done with instruction editing or inpainting, which target a region. Plain image-to-image tends to move everything a little.