Find the conflicting instruction
A fictional team submits aspect_ratio: "4:5" and resolution: "4k" but the text prompt asks for a square 2K portrait. FASHN Agent’s experimental API explicitly prioritizes a ratio or size requested in the prompt over those structured fields. The conflict is therefore a request-authority problem, not necessarily an export defect. Inspect template fragments and prior instructions as well as the visible JSON; do not assume the structured fields always win.
Use a precedence matrix
If the prompt says square while the field says 4:5, reconcile the prompt rather than adding more field copies. If no ratio is stated in the prompt, the structured aspect_ratio defines the requested shape. If neither specifies one, the Agent generally keeps the frame when editing and chooses a suitable frame for a new image. Apply the same conflict check to output size. This behavior is specific to Agent, not a general rule for every FASHN endpoint.
| Condition or role | Decision or evidence | Next check |
|---|---|---|
| Prompt square; field 4:5 | Prompt wins | Remove contradiction |
| Field 4:5; prompt silent | Structured ratio applies | Inspect actual output |
| Neither states ratio | Agent chooses or retains frame | Specify destination shape |
| Prompt 2K; field 4K | Prompt size wins | Resolve size conflict |
Separate tier from returned pixels
The provider notes that some operations can be made and billed at 2K even when 3K or 4K was requested. After an authorized run, inspect each result’s actual pixel dimensions and its prediction’s credits. A requested 4K field is not proof of native 4K delivery. In the fictional catalog, the team separates a 4:5 listing tile from a 3:2 hero image into explicit deliverable requirements, removing contradictory ratio words before another run. No API output or invoice was inspected.
Diagnose the right failure
If the ratio is wrong, inspect text-versus-field precedence. If the ratio is correct but the image is too small, inspect actual output dimensions and operation tier. If the crop is poor, that is a composition issue. The Model Create face-reference cap is a different endpoint condition, while Reframe’s already-matching ratio error is a separate validation mechanism. Record exact prompt, fields, returned dimensions and credits; hold any file below the approved native-dimension requirement.