Text-to-image generates a picture from a written description.
The prompt is embedded by a text encoder such as CLIP.
That embedding steers the denoising toward matching content.
Guidance scale controls how strictly it follows the prompt.
Too much guidance produces rigid, over-saturated, unnatural images.
Text-to-image turns a written description into a picture.
The bridge between the two is a shared representation of text and images.
CLIP built that by training on hundreds of millions of image and caption pairs.
A text encoder turns the prompt into vectors in that shared space.
Those vectors condition the denoising at every step through cross-attention.
So the prompt is not consulted once but continuously as the image forms.
Guidance scale decides how hard the prompt pulls.
Classifier-free guidance runs the model with and without the prompt and amplifies the difference.
Around seven is typical, and above twelve images turn rigid and over-saturated.
Negative prompts describe what to avoid and are subtracted the same way.
Prompts work best when concrete, since a specific noun beats an adjective.
Style, medium, lighting and composition are the levers that change results most.
Known weaknesses persist: hands, text within images, and counting objects.
Spatial relationships are unreliable, since a red cube on a blue sphere often comes out reversed.
The training data was scraped, which is why copyright and consent remain live questions.
One question on this card. Every option explains itself once you pick, right or wrong.
18 concepts in this module, then a graded test with every answer explained.