“Use Image 1 for the actor” is only useful when Image 1 actually refers to the actor file in the current request. Reordering, replacing or omitting an attachment can make an unchanged prompt point to the wrong source. Audit the physical attachment list before adjusting role priorities or judging the model’s fidelity.
Establish the received attachment order
Jimeng’s official image-generation introduction illustrates numbered references for different purposes. Seedance’s official prompt guide discusses reference dimensions such as appearance or motion. These product-specific examples do not prove that another interface interprets “@Image 1” identically. Use the current supported input convention.
Number the actual individual attachments in their submitted order. For each, record a non-sensitive filename or internal identifier, media type, recognizable thumbnail description and intended use. Where a request manifest or task receipt exposes received files, compare it with the local list. If that evidence is unavailable, state that limitation rather than asserting server-side binding from the prompt text alone.
This is a file-to-label audit, not a contact-sheet panel audit. A single contact sheet with six internal panels remains one attached file unless the interface explicitly supports panel addressing. Likewise, a reference-priority rule cannot repair a label that points to the wrong uploaded object.
Test a deliberate two-file reorder
Use an actor photograph and a separate background image in this proposed diagnostic plan. Begin with actor then background; swap their upload order in a saved draft. Before any generation, update the labels from the actual list and compare each displayed thumbnail with its purpose.
| Label after reorder | Actual attached content | Allowed reference purpose |
|---|---|---|
| Image 1 | Background file | Scene layout and environment |
| Image 2 | Actor file | Main character appearance |
- Preserve the old list and prompt as a distinct version.
- Record the new attachment order and recognizable thumbnails.
- Rewrite only the labels affected by that order change.
- Compare the submitted manifest if the interface exposes one.
- Save the label-purpose mapping with the request and result identifier.
The table is an example to fill with real evidence, not an executed binding test. Do not silently carry over numbering from an earlier conversation when the current interface presents a different list.
Separate input acceptance from result acceptance
The input passes when each supported label maps to the intended actual attachment and its purpose is explicit. The generated output then needs an independent review of character, scene and motion. An accurate manifest does not guarantee that the model followed it throughout a clip.
Imagild Skills can help prepare visual references; the external generation system owns its file-binding convention. Include the exact mapping and any missing receipt evidence in the handoff. When the output is wrong, this record lets an editor distinguish an attachment-order mistake from a model-following failure instead of repeatedly rewriting unrelated prompt priorities.