Composite Scene merges multiple real images into one coherent AI-generated scene by reconciling placement, scale, lighting, and perspective.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install composite-scene
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install composite-scene using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Composite Scene?
Openclaw Skills Composite Scene is a multi-reference image workflow for fusing separate real images into one believable composition without manual cut-outs, masking, or hand compositing. It is designed for prompts like put this product into that scene or combine these two photos, where each source image contributes a distinct element and the model handles relighting and perspective correction.
This skill is the reverse of character-consistency: instead of preserving one subject across many images, it brings many images into a single scene. It works best when you provide one clean reference per element, a clear target scene, and explicit instructions for placement, contact, scale, and lighting.
Composite Scene Use Cases
- E-commerce and ad creatives where a clean product shot needs to be dropped into a styled environment.
- Subject-and-background composites such as
place my subject on this backgroundordrop the watch onto the table. - Multi-photo merges where two or more people, objects, or scenes must read as one coherent frame.
- Style-transfer composites that pair a content reference with a style reference for a repaint in the target look.
- Batch scene variants where the same references are reused across multiple environments or layouts.
How Composite Scene Works
- Resolve the live model schema with
runware-runand confirm the reference-image field, allowed count, and exact parameter names. - Collect one clean reference per element and upload each image as its own entry in
inputs.referenceImages. - Write a
positivePromptthat names each element by position (the watch from the first image,the table from the second image) and states the relationship, scale, contact, camera angle, and target lighting. - Run
imageInferencesynchronously with the live compositing model, settingwidth/heightorresolutionand aseedif you need reproducibility. - Review the composite for identity, placement, shadows, and perspective. If the scene is busy, feed a strong partial result back as a new reference and add the next element in a second pass.
- Reuse the same references for alternate scenes by changing only the scene clause, then vary
seedto explore controlled variations.
Composite Scene Setup
- Confirm the current live model before you build prompts. Openclaw Skills should not be tied to a stale model choice.
- Inspect the live schema so you know the exact reference-image field name and maximum reference count.
- Prepare clean source images: one image per element, preferably with simple backgrounds for predictable compositing.
- Send the references as separate array entries, then direct placement and lighting in the prompt.
# Inspect live image models and confirm the current composite-capable option
runware-models
# Resolve the live schema before sending the compositing request
runware-run
- If you are composing a complex scene, build it in passes rather than stacking every reference in one shot.
- Save the final composite, and reuse the same references when you need more scene variations.
Composite Scene Data Schema & Taxonomy
Generated artifacts:
- Final composite image for each run.
- Optional intermediate composites if you choose to save a partial result and feed it back in.
- No manual masks, cut-out layers, or compositing files are required.
| Field | Type | Purpose | Notes |
|---|---|---|---|
inputs.referenceImages |
array of image references | Supplies one source image per element | Order matters; the prompt refers to the first image, the second image, etc. |
positivePrompt |
string | Controls placement, scale, contact, lighting, and relationship | This is the primary compositing control surface. |
seed |
integer | Reproduces or varies a result | Fix it for iteration, change it for alternates. |
width / height or resolution |
dimensions | Defines the output frame | Match the target scene framing. |
model |
live model identifier | Selects the compositing-capable image model | Default guidance points to Google Nano Banana 2 (google:4@3) when live. |
Metadata taxonomy:
- Identity: what each reference element is.
- Position: where the model should place each element.
- Relationship: how elements touch or relate (
resting on,leaning against,walking beside). - Lighting: the single target light setup to reconcile across the frame.
- Camera: angle, framing, and perspective for the final scene.
- Iteration state: the current pass, reused references, and chosen
seedfor repeatability.
Composite Scene Advanced Features
- Supports up to 14 reference images in one call on Google Nano Banana 2, making Openclaw Skills strong for dense multi-element composites.
- Uses live schema discovery through
runware-modelsandrunware-run, so you can confirm current field names and limits before every request. - Lets you reuse the same reference set across multiple scenes, which is ideal for product variant generation and batch creative testing.
- Handles cross-element relighting and perspective reconciliation so objects do not look pasted in.
- Works in passes: confirm a partial composite, then feed it back in as a new reference to add complexity safely.
- Supports reference-guided alternatives such as Nano Banana Pro and IP-Adapter on FLUX/SDXL when the default model is unavailable.
- Preserves honest composites by keeping source identities intact and avoiding invented logos, badges, or copy that was never supplied.
SKILL.md
Loading
Related Openclaw Skills
Controlled Generation (ControlNet)
Preserve an image’s structure while changing its style, subject, or finish with ControlNet-driven generation.
2D Game Assets
Openclaw Skills for 2D game assets generates cohesive sprite sets, icons, stickers, and transparent cutouts that all share one visual style.
Image to 3D Asset
Openclaw Skills converts photos or prompts into production-ready GLB 3D assets with controllable topology, materials, and polygon budgets.
LLM Agent
Build OpenAI-compatible LLM agents on Runware that can reason, call your functions, and answer with grounded tool results.
Cinematic Video
Generate short, shot-directed videos that feel filmed rather than synthesized, using precise cinematic prompts and live model-aware workflows.
Learning English from TV Series (DramaLex)
DramaLex turns a TV episode or film into a complete English-learning loop with subtitle mining, CEFR calibration, drills, and review.
Featured*
