Text to image
Text to Image AI
Describe the picture. Imagesmith sends that sentence to the model you pick and shows the image here. Nothing about the request is stored on this site.
Your image will show up here.
From noise to a picture, one step at a time
Text-to-image models are usually diffusion models. In training they see real photos get steadily scrambled, and they practice guessing the scramble so they can undo it. Making a picture runs that backward. The frame begins as random noise. The model then takes a short series of small steps, each one clearing a little of that noise, with your prompt as the guide. The slider below walks through a sketch of those steps. It is drawn here in the browser, not returned by the generator above.
Signal 0% · noise 100% (cosine schedule)
Early steps are still mostly static. Only the largest masses, such as sky against ground, start to separate.
The same path explains three things that surprise people. Each run begins from a different noise field, so the same prompt does not repeat itself. The layout is chosen while the frame is still soft, which is why a prompt that places the subject (“fox on the left, open snow on the right”) holds up better than editing the arrangement later. Thin details, including letters and fingers, arrive only at the end, and that is where small errors usually remain.