Learning to remove noise
In a basic training setup, different amounts of noise are added to images. The model learns to estimate the noise, or a related quantity, so it can guide the recovery of structure.
Generation is not a search for a finished picture in a folder. The model uses learned patterns to gradually change its current representation of an image.
From randomness to structure
During generation, a text description can guide these changes. In many systems, the process takes place in a compressed feature space, before a separate decoder turns the result into pixels.
Why results differ
The initial random state, computational settings and the model itself affect the result. The same phrase can lead to different images. The diagram above illustrates the stages; it does not generate an image.
KEEP EXPLORINGMake sense of another idea↗