Made by AI, Directed by David Stein

Search
AI & Creative Tools

Img2Img: Getting Started

The simplest way to steer an AI image model with something more specific than a text prompt — a starting image.

3 min read

Every text-to-image generation starts from random noise and a prompt. Img2img starts from a real image instead — a photo, a sketch, a previous generation — and nudges it toward whatever your prompt describes. It’s the single most useful technique to learn after basic prompting, because it answers the question text-to-image can’t: how do I get a result that’s close to this specific thing, not just close to these words?

How it actually works

Diffusion models generate images by starting from pure noise and removing a little bit of that noise at each step, guided by your prompt, until a coherent image is left. Img2img changes the starting point: instead of pure noise, it starts from your source image with some noise added back in, then runs the same denoising process from there.

Denoising strength (sometimes called “strength” or “noise”) controls how much noise gets added before denoising starts — and therefore how much of the original image survives the process:

  • 0.1–0.3: light touch. Composition, pose, and most details stay intact. Good for style transfer, color grading, or small texture changes.
  • 0.4–0.6: moderate changes. Composition generally holds, but details, lighting, and surface texture can shift substantially.
  • 0.7–0.9: heavy changes. The source image becomes more of a loose compositional suggestion than a thing being edited — useful for using a rough sketch as a starting point, not for preserving a photo.
  • 1.0: equivalent to text-to-image. The source image contributes essentially nothing.

The mistake almost everyone makes first

Img2img cannot fix a composition problem with a low denoising strength. If a generated image has the wrong pose, the wrong camera angle, or an object in the wrong place, running it back through img2img at 0.3 strength won’t fix it — low strength preserves exactly the structure you’re trying to change. Fixing structural problems requires either a much higher strength (which risks losing everything else you liked) or a different tool entirely, like ControlNet, which lets you lock in a pose or depth map explicitly rather than hoping a strength value threads the needle.

The useful mental model: denoising strength trades fidelity to the source for freedom to fix things, on a single dial, with no way to get both at once. If you need to keep the composition exactly and change something structural within it, you need a more targeted tool.

Common uses, by strength range

  • Upscaling with added detail (0.2–0.4): generate at a lower resolution, then upscale and run img2img at low strength to add fine detail without changing the image.
  • Style transfer (0.3–0.5): take a photo, apply a painterly or stylized prompt, keep the subject recognizable.
  • Sketch-to-render (0.6–0.8): take a rough drawing or 3D render and let the model add realistic lighting, texture, and detail while following your rough shapes.
  • Variation generation (0.5–0.7): take a generation you like and produce several variations that are clearly related but not identical.

A practical workflow

  1. Start at 0.5 strength as a baseline and generate one result.
  2. If too much of the source survived (the “fix” barely shows), raise strength in increments of 0.1.
  3. If too little of the source survived (it looks like an unrelated new image), lower strength instead.
  4. Once strength is roughly right, refine your prompt rather than the strength value — prompt changes are more predictable once you’re in the right strength range.

Img2img is a dial, not a feature — understanding what it trades off is most of what there is to know.

A walkthrough of denoising strength and common img2img workflows.

Related Reading