Style Transfer via Vision Model Description

Use a vision model (e.g., Qwen3-VL-2B) to read a reference image and output a text description of its style (e.g., '2D cartoon illustration, bold outlines, flat colour areas, warm earthy tones, painterly brushstrokes'). Then use an image generation/editing model (e.g., Qwen-Image-Edit 2509) with that text description to render a new scene or repaint an existing image, preserving the style without transferring objects or people. This works locally with tools like ComfyUI, using 8 steps with a Lightning LoRA (~35 seconds on a 4090).