An editor’s ability to import many photographs into a timeline does not establish how many references an AI feature receives in one request. These are different input contracts. Identify the active tool and model, read its current constraints, and confirm the accepted attachment set before generation. Do not infer receipt from the number of files you selected initially.
Separate library, timeline and model inputs
CapCut’s batch-image help distinguishes ordinary editing from AI inputs. Volcengine’s Jimeng image API documentation describes a separate supplier interface. Neither source defines a universal allowance for Jianying, every model or Imagild. Limits must be read from the official documentation and current interface for the exact active workflow.
List four scopes: files stored in the library, media placed on the timeline, attachments accepted in the present AI request, and any documented conversation-wide constraints. A picture stored in the first scope has not necessarily entered the third. A conversation can also contain older attachments that are not clearly assigned to the current request.
Do not publish a generic “nine pictures” or another fixed number as a rule across products. Model versions, modes, file formats and endpoints can differ. Record the check date and the relevant current constraint when preparing a real request.
Build a received-input manifest
For a proposed test, prepare six references with independent purposes: subject appearance, clothing detail, backdrop, object, palette and composition. Before submitting, inspect which files the interface actually accepts. If only four are received, adjust the plan instead of continuing to refer to two missing images as if the model saw them.
- Identify the active feature, model and request mode.
- Read its image count, type, byte-size and dimensional constraints separately.
- Match the accepted attachments to the six planned purposes.
- Resolve rejected or omitted inputs before submitting a paid generation.
- Save the final manifest with the prompt and result identifier.
Splitting references across messages is not automatically equivalent to one multimodal request. Use that route only when the product documents how the active context handles earlier attachments. When the contract is unclear, reduce the request to a testable supported set rather than inventing persistent memory behavior.
Check outputs against the accepted set
Acceptance at the input stage means every required reference has an identifiable received attachment or has been deliberately removed from the plan. It does not mean the model followed that reference successfully. After delivery, compare the actual result with the manifest and record unsupported or missed requirements separately.
This is an authored test plan, not a successful six-image experiment. A rejected file is an input problem; an accepted background that is ignored in the output is a result problem. Keeping those distinctions prevents prompt rewriting from disguising a missing attachment. Use Imagild Skills only with the inputs that the selected workflow actually supports; external model documentation is not a promise that the same multimodal feature exists there.