Automatic schema detection

Replicate models disagree on how they take an image size. Some accept pixels, some only a fixed list of aspect ratios, some both. imgen does not keep a list of models for this. It asks the API.

What imgen reads

Before the first prediction, imgen fetches GET /v1/models/<owner>/<name> and looks at the OpenAPI schema of the model’s latest version:

1. aspect_ratio under components/schemas/Input/properties. The allowed values come from an enum on the property itself or, if the property points to a shared definition via $ref (also inside allOf), from that definition.
2. output_format in the same place. If it exists, the model can deliver PNG or WebP itself.

The result is printed once per model:

Resolving model black-forest-labs/flux-1.1-pro…
  version: <first 12 characters of the version id>…  aspect_ratio: <allowed values or none>  output_format: <yes|no>

How the size is sent

The schema has …	imgen sends
aspect_ratio including custom	aspect_ratio: "custom" plus your exact width and height
aspect_ratio with fixed values only	the allowed ratio closest to width / height, no pixel size
no aspect_ratio	width and height as given

“Closest” compares the numeric ratios. With the default 1024×768 (1.33) and a model that allows 1:1, 16:9, 3:2 and 4:3, imgen picks 4:3. For 1920×1080 it picks 16:9. If none of the allowed values has the form W:H, it falls back to 1:1.

With fixed ratios, the model decides the final pixel size. Your --width and --height only pick the shape.

How the format is sent

If the schema has output_format, imgen requests png, or webp with output_quality when WebP output is enabled. Otherwise it downloads what the model returns and converts locally if needed. Details in WebP output.

Why this matters

Before version 0.3.0, imgen only read a direct enum. FLUX Pro defines its ratios through $ref, so every FLUX Pro image came out square. Following the reference fixed that without any model-specific code.