Image decomposer

Image Decomposer

Decompose any composite image into individual layers and elements. Takes a screenshot, banner, ad, mockup, or game asset and creates a workflow that extracts each element in isolation.

Phase 1: Image Analysis

Analyze the provided image using vision. Identify and list all distinct visual elements with their categories:

Category	Aspect Ratio	Remove BG	Description
Background	Original ratio	No	Scenes, landscapes, ambient backgrounds, effects, patterns, textures
Character	1:1	Yes	People, mascots, avatars, animals
Icon	1:1	Yes	Small symbols, emojis, badges
Text/Title	1:1	Yes	Headings, stylized typography
UI/CTA	1:1	Yes	Buttons, timers, progress bars, badges
Logo/Brand	1:1	Yes	Logotypes, watermarks, brand marks
Product	1:1	Yes	Packshots, physical objects, merchandise
Decorative	1:1	Yes	Frames, borders, ornaments, sparkles, ribbons

Present the detected elements as a numbered list:

Detected elements:
1. Pirate character (Character) - 1:1 + RemoveBG
2. Ship deck and ocean background (Background) - 16:9
3. Title "Plunder the Loot" with skull decorations (Text/Title) - 1:1 + RemoveBG
4. Reward icons: gold coins, cash bills (Icon) - 1:1 + RemoveBG
5. Ship wheel with target board (Decorative) - 1:1 + RemoveBG
6. CTA button "Default" (UI/CTA) - 1:1 + RemoveBG

Phase 2: User Confirmation

Let the user review and adjust the element list:

• Remove elements: "remove #6"
• Add elements: "add the corner skulls as decorative"
• Edit categories: "#5 is an icon, not decorative"
• Merge elements: "merge #4 and #5"
• Confirm: "ok" / "looks good"

Phase 3: Build Workflow

Use xainflow_create_workflow to create a workflow named "Decompose - [image name]".

Workflow Structure

Build the workflow with these nodes:

Asset node (original image):

Position: (0, 0)
Type: asset
Data: { selectedAsset: { file_url: "[IMAGE_URL]" }, output: "[IMAGE_URL]" }

For each element, create an imageGenerator node:

Position: (400, element_index * 200)
Type: imageGenerator
Model: gpt-image-1-5 (or user's choice — models with high reference support work best)
Ratio: category ratio from table above

Extraction prompt for each element:

Extract only the [element description] from this reference image.
Generate it isolated on a clean white background.
Maintain the exact same style, colors, proportions, and details as in the original.
Do not add, remove, or modify any details.
Output only the [element description], nothing else.

For elements that need background removal, add a removeBackground node:

Position: (800, element_index * 200)
Type: removeBackground

Edges

• Asset → each imageGenerator (asset-input)
• imageGenerator → removeBackground (where applicable)

Layout

Nodes organized in 3 columns:

Col 1 (x=0)          Col 2 (x=400)              Col 3 (x=800)
[Asset: Original]  -> [Gen: Background]
                   -> [Gen: Character]         -> [RemoveBG: Character]
                   -> [Gen: Title]             -> [RemoveBG: Title]
                   -> [Gen: Icons]             -> [RemoveBG: Icons]
                   -> [Gen: Decorative]        -> [RemoveBG: Decorative]
                   -> [Gen: CTA]               -> [RemoveBG: CTA]

Vertical spacing: ~200px between nodes.

Recommended Models

Models with high reference support for faithful extraction:

• GPT Image 1.5 (16 references) — highest fidelity
• Nano Banana Pro (14 references) — good balance
• SeeDream 4.5 (10 references) — alternative

Before Executing

Show a summary with:

• Number of elements to extract
• Which elements get background removal
• Model being used
• Estimated credits: ~[N elements] x [model cost] + [M removals] x 2 cr

Ask for confirmation.

Edge Cases

• Very complex images (20+ elements): Group similar items (e.g., "all reward icons" as one extraction)
• Abstract/artistic images: Propose broad categories, let user refine
• Text-heavy images: Each text block can be separate or grouped