Using multiple references: separate subject, composition and style

Give each reference one main role

Start with the subject image, add composition or style references only as needed, and describe their separate roles in the prompt. Before submission, use the actual items and order in the Composer input bar. Grouping images does not assign their roles for you.

A product photograph supplies appearance, a sketch supplies spatial relationships, and a style image supplies lighting or mood. If several images compete to define the subject, the model may mix people, packaging or clothing. Start with essential references and replace blurred, heavily occluded or structurally incomplete images.

Check what the model accepts

• Reference counts, accepted media types and required inputs depend on the current model and workflow. One model's limits do not apply to every model.
• Image imports must satisfy both the 20,250,000-pixel total limit and the 8192px longest-edge limit. If exceeded, choose to resize for upload or cancel, then check small text and subject details after resizing.
• After upload, confirm the asset appears in the current input bar. Placing a file on the canvas, writing its local filename or entering a folder path does not replace adding the reference.

Prepare task inputs from canvas groups

1. Group the subject, composition and style assets for the same task.
2. Select the group, or box-select it together with other assets, then inspect the individual members listed in Composer.
3. Remove unwanted inputs individually without dissolving the group or deleting the source assets.
4. Describe image roles using the current input order. Check numbering again after removing or adding an item.
5. Check the model, input count and parameters before submitting. Canvas filters change what is displayed; they do not replace checking the input bar.

Workflow canvas picking fills one parameter

When you start canvas picking from an image parameter in an extension workflow, the chosen image fills only that initiating parameter. It does not replace the entire Composer input set or change canvas group membership or order. Pick individual members when browsing a group.

Back in the form, check each main-image and reference-image slot. Cancelling preserves the parameter's previous value. Unfinished, expired or wrong-type assets cannot be committed as that pick.

Author-written multi-image prompt, not generation-tested

This three-image example requires a model that currently accepts at least three images. It has not been tested in a generation run. Check input order first. If only one image is supported, retain the subject reference and describe composition and lighting in text.

Image 1 defines the product appearance; preserve its shape and logo. Image 2 supplies only the top-down composition. Image 3 supplies only soft cool lighting; do not copy its people or text. Keep the entire product visible and add no other brands.

Common errors: mixed subjects, rejected inputs or wrong slots

Review identity, shape, logos and packaging text before judging style. Keep the initial subject image and approved results, so repeated generations do not become the only reference and accumulate drift.

• Mixed subjects: return to the subject image alone, then add composition and style references one at a time.
• Style overrides product appearance: limit the style reference to lighting or background, and remove references containing a conflicting subject.
• Too many inputs or an unsupported media type: remove incompatible items from the input bar or select a workflow that actually supports them.
• A parameter is still wrong after picking an image: check whether you used Composer inputs or the picker for a specific workflow parameter.