Diffusion model: what it is

A diffusion model is a family of generative AI techniques used to create and edit images. The name comes from the process: during training, the model learns how images fall apart when noise is added little by little, and then learns to do the reverse, removing noise in steps until a sharp image forms. In generation, that process is guided by an instruction (a text, for example) or by a starting image.

How does this become photo editing?

To edit a photo, the model receives the original image and an instruction about what to change, and redraws the image keeping what should stay and altering what was asked. It is the idea of inpainting, in which a region is redrawn consistently with the rest. Quality depends on the model, the instruction and the input image.

Because the process starts from probabilities, two attempts with the same input can give slightly different results. This is normal and should be taken into account when interpreting any simulation.

What are the limits?

Understanding the limit is part of responsible use:

• the model produces a plausible image, not a measurement or a prediction of a clinical result;
• it can "invent" details that were not in the photo, a phenomenon called hallucination in image AI;
• it is sensitive to photo quality: poor light, angle and low resolution worsen consistency;
• it does not know the product, the technique or the patient's biological response.

For that reason, any AI-generated image shown to a patient must be labeled as illustrative. See also AI image provenance.

Why does an HOF professional need to know this?

Because the patient will ask: "is this how it will look?". Understanding the logic of the model makes it possible to answer honestly: the image is a coherent illustration of what was requested, useful for talking about expectations, but it is not a promise. When the professional understands the tool, they use it better and explain it better.

In Harmy

The app sends the frontal photo, with the professional's consent, to a third-party AI on the server to generate the projection. The prompt respects the volume in ml entered and forbids makeup, skin retouching, and changes to expression, background, hair and lighting, so that only what was marked changes. The image is always an illustrative simulation, not a promise of results.

Also called: diffusion, diffusion model.