When an output becomes square or portrait after reference reordering, inspect the default canvas rule before concluding that the model ignored your composition image. Alibaba’s Wan image API documents a last-input aspect-ratio rule for relevant default size tiers. Apply that rule only to the named model and task configuration, not to all multi-image generators.
Record the canvas separately from reference roles
Create a list in the actual submitted order. For every picture, record its dimensions and intended role: identity, garment, product detail or composition. Then record whether output dimensions were explicitly selected or left at a default tier. The order visible in a planning document may differ from the order actually sent.
Imagine a hypothetical request containing a horizontal room photograph and a vertical portrait. You want the portrait identity inside a horizontal room. If the portrait is the final image and the task derives its default shape from that position, the returned portrait canvas need not mean the room reference was rejected. It can be a canvas decision made before any meaningful content review.
Make the output requirement explicit where supported
Check the exact API or interface for supported size controls. Set a permitted explicit canvas when that is appropriate. If only a documented default relationship is available, arrange inputs deliberately and retain the role labels. Do not reorder images blindly: changing order can also change how references are labeled or interpreted.
| Submitted position | Image dimensions | Visual responsibility | Canvas implication |
|---|---|---|---|
| First | Recorded from file | Environment | Does not necessarily determine default shape |
| Last | Recorded from file | Person identity | Check documented default rule |
| Output setting | Exact value or default | Delivery canvas | Verify actual downloaded dimensions |
This table is an original request audit. It does not claim that a particular image order was tested or that explicit sizing guarantees composition fidelity.
Inspect shape before inspecting content
Read the downloaded image’s actual width and height. Separate three questions: Is the canvas correct? Is the subject placed correctly? Did the required identity or product detail survive? A horizontal file can still crop away the subject, while an incorrect portrait file can contain a faithful face. Record those as different verdicts.
Do not fix a wrong canvas by stretching the image. Consider an appropriate crop, expansion or new supported request, depending on how much content is missing. A crop cannot recover an environment never generated. If changing the canvas alters protected content, return to a reviewed source rather than carrying an unapproved face forward.
Preserve a reproducible handoff
Save the actual attachment order, reference labels, size setting and output dimensions together. If a collaborator adds another detail reference, check whether it becomes the new final input and changes the default behavior. Keep canvas decisions visible instead of relying on a remembered upload order.
Imagild image-editing work can use this record to state the required deliverable, but Wan’s default rule does not establish the same rule in Imagild. Check the selected workflow’s actual controls and review the returned file independently.