Compose Multiple Images Into One Scene

2–9 reference images → one unified, believable composite scene.

Category: edit · Difficulty: Advanced · ~Time: 5 min

What you'll make

Merge several inputs — a person, a product, a location, a style reference — into a single coherent image where lighting, scale, and perspective all agree. Use it to place your subject in a new setting, build a group shot from separate portraits, or stage a product in a curated scene.

Models used

🎨 Image — Seedream 4.5 (sequential) (resolve live; e.g. bytedance/seedream-v4.5/sequential)
🎨 Image (alt) — Nano Banana Pro (multi-ref edit) (resolve live; e.g. google/nano-banana-pro/edit)
💬 LLM — Claude (to write the composition / staging brief), optional

Steps

Upload every reference — "Upload each of these local images and list their URLs." → maps to atlas_upload_media once per file; collect all download_urls.
(Optional) Draft a staging brief — "Write a one-paragraph scene brief placing the subject from image 1 into the location from image 2." → maps to atlas_chat with Claude.
Compose — "Combine these references into one scene: keep the person from ref A, the product from ref B, place them in the setting from ref C, match lighting and scale." → maps to atlas_generate_image with Seedream 4.5 (sequential), passing the uploaded URLs as the reference/input images.

Tip: ask the agent to atlas_search_docs for the model first — the multi-image input field name and the max number of refs vary by model, so pull the live schema.
Unify — if the composite looks pasted, re-run asking for "a single consistent light source, matched color temperature, and correct relative scale."

Prompt starters

Compose one photorealistic scene from the references: the woman from image 1,
wearing the jacket from image 2, standing in the cafe from image 3. Match a
single warm light source, consistent perspective, and realistic relative scale.

Place the product from image 1 onto the marble countertop from image 2 in the
style/mood of image 3. Unified lighting, accurate reflections, one cohesive shot.

Tips & variations

Seedream 4.5 (sequential) is built for multi-reference composition; Nano Banana Pro (edit) is a strong multi-ref alt with excellent subject consistency.
Number your refs in the prompt ("image 1", "image 2") and state each one's role — it dramatically improves which element the model keeps from where.
Fewer, cleaner refs beat many noisy ones. Crop each input to just the element you want before uploading.