Image to image
Starting from a picture instead of pure noise, and redrawing it in part.
In plain words
In image to image the sampler does not start from an empty latent. You load a picture, encode it into a latent, add some noise, and let the model remove that noise its own way. With denoise you choose how much of the picture is redrawn: little noise keeps the composition and changes the surface; much noise keeps only the rough layout.
Uses: change the style of a photo or sketch, add detail to a rough drawing, refine a picture with another model, make variations of a picture you like.
Where you meet it in a graph
Load Image (IMAGE) → VAE Encode (pixels, vae) → KSampler's latent_image, with denoise below 1. The rest is the text to image graph. Often an Upscale Image or ImageScale node sits before VAE Encode to bring the picture to the model's size.
Editing models are a different way to change a picture with words: Qwen-Image-Edit, FLUX.1 Kontext and Flux.2 read the picture as a reference and follow an instruction ("make the coat red"), often with denoise 1. They have their own templates; see Reference images.
Values and what they change
- denoise 0.3: same picture, new texture; 0.5: same layout, new details and style; 0.7 and above: a new picture loosely based on the old.
- The picture's size decides the latent's size: resize huge photos first.
- Write the prompt for the result you want, not for the source.
Typical mistakes
- denoise left at 1: the source is ignored.
- A picture far larger than the family's size: slow, repeated objects.
- Encoding with one model's VAE and sampling with another family.
How the app helps
The lesson A picture from a picture builds it; Refining with a second model uses it as its second half. The VAE Encode description explains the start.