How AI photo merging works
Reference images provide visual information, while your prompt describes the intended scene. The result is newly generated; it may reinterpret faces, objects and backgrounds rather than faithfully copying every pixel.
Choose suitable source photos
Prefer clear subjects, compatible lighting and simple backgrounds. Two well-lit portraits are easier to interpret than a collection of small, obscured faces. Decide which photo contributes each subject or setting before uploading.
Prepare your images
Rotate a sideways source and crop away unrelated objects. Keep enough surrounding space for the subject to remain recognizable. Apply the edit before generating so the adjusted file, not the initial upload, is submitted.
Write composition instructions
Describe every important subject and its placement. For example: ‘The person from photo one stands on the left; the person from photo two stands on the right. Put them together in a garden under soft afternoon light.’ Ask for one cohesive scene rather than a collage.
Choose strength and canvas orientation
Begin with moderate strength and select a landscape canvas when people need space side by side. The current generation service uses three size groups; inspect the output dimensions before planning a precise layout.
Review and improve
Check faces, hands, edges and the relationship between subjects. If one person is missing, explicitly identify that person in a revised prompt. If the image is overly stylized, reduce strength. Change one instruction at a time so you can understand its effect.
Common mistakes
Too many references, contradictory prompts and poor source lighting can complicate a result. Asking for perfect identity preservation or accurate small text does not ensure it. Treat a generation as a draft that needs review.
Final checklist
Before sharing, confirm that the image communicates your intended idea and that you have permission to use the sources.
- Every intended subject is visible
- Hands, faces and object edges have been reviewed
- The image is not presented as an unedited real event
- The downloaded dimensions fit the destination
- Source-image rights and privacy have been considered
