The shape of a picture, in words every image model can be given.
Providers disagree on how a shape is asked for -- OpenAI takes a pixel size from
a short list, Fal an aspect_ratio, Together a width and height -- so this is
deliberately not any of them. Each image_models.json entry translates it into
its own arguments, along with what that shape costs, because a larger canvas is
more tokens on a token-billed model.
'square': 1:1, and what every entry draws when nothing asks otherwise.
'landscape': wider than tall, 3:2 where the model offers it.
'portrait': taller than wide, 2:3. A shape, unrelated to character portraits.
Asking for it in the prompt does not work: the canvas is an API argument, and a
"wide shot" requested of a square canvas comes back square.
The shape of a picture, in words every image model can be given.
Providers disagree on how a shape is asked for -- OpenAI takes a pixel size from a short list, Fal an
aspect_ratio, Together a width and height -- so this is deliberately not any of them. Eachimage_models.jsonentry translates it into its own arguments, along with what that shape costs, because a larger canvas is more tokens on a token-billed model.Asking for it in the prompt does not work: the canvas is an API argument, and a "wide shot" requested of a square canvas comes back square.