Autonomous Spatial Apparel Placement via Multimodal Vision Pipelines

import os
from google import genai
from PIL import Image

client = genai.Client()

def generate_folded_apparel_prompt(sku: str, color: str, location: str) -> str:
    """
    Generates deterministic multimodal prompt for retail planogram rendering
    """
    return f"""Generate a high-resolution, realistic product image of a folded {color} garment 
(SKU: {sku}) for display at {location}. The final image should meet the following criteria:

1. Folding Style: Neatly folded as in retail packaging; maintain symmetry and alignment in folds.
2. Realism & Texture: Natural fabric texture (cotton, linen, denim); soft diffused daylight only.
3. Framing: Centered 1024x1024 px square output with clean spacing.
4. Background: Pure white or transparent background suitable for planogram ingestion.
5. Exclusions: No human models, mannequins, props, dramatic lighting, or stylized edits.
6. Color Fidelity: Exact {color} fidelity without tinting or filters."""

from PIL import Image
import numpy as np

def align_garment_to_canvas(image: Image.Image, target_size=(1024, 1024)) -> Image.Image:
    """
    Calculates alpha bounding box and centers garment on target transparent canvas
    """
    if image.mode != 'RGBA':
        image = image.convert('RGBA')
    
    alpha = image.split()[-1]
    bbox = alpha.getbbox()
    if bbox is None:
        return Image.new('RGBA', target_size, (0, 0, 0, 0))
    
    cropped = image.crop(bbox)
    scale_factor = min(target_size[0] / cropped.width, target_size[1] / cropped.height) * 0.9
    new_size = (int(cropped.width * scale_factor), int(cropped.height * scale_factor))
    scaled = cropped.resize(new_size, Image.Resampling.LANCZOS)
    
    result = Image.new('RGBA', target_size, (0, 0, 0, 0))
    x = (target_size[0] - new_size[0]) // 2
    y = (target_size[1] - new_size[1]) // 2
    result.paste(scaled, (x, y), scaled)
    return result