E-commerce-Grade Product Image Quality: Reference-Image-Driven Still Life Retouching

Using an ordinary product still-life photo as the reference image, redo the photographic quality with the gpt-image-2.5 Sunburst image editing pipeline: keep the product's shape, labels, and proportions, and only re-light, re-set the background, and redo the material highlights, outputting a finished image that can be used directly for e-commerce detail pages and advertising assets. In terms of composition, base only provides resolution (2K, balancing detail and time cost) and aspect_ratio (1:1 square, fitting most e-commerce main-image slots), and does not enable background (transparent backgrounds only support 1K, and a high-precision opaque background is needed here). The reference image is generated from one image asset in assets and bound to image_urls[0]; it serves as both the "input image" and the "editing anchor"; the top-level prompt is automatically written into input.prompt by the pipeline. The prompt approach is organized in four layers: subject lock (shape, color, printed text, and proportions unchanged), lighting (large softbox + single-side rim light + controllable tabletop reflection), materials (brushed metal, matte plastic, and glass translucency each described in terms of roughness and highlight form), and composition and post-processing (macro depth of field, clean background, and negative space to accommodate copy). Applicable scenarios: batch re-creation of e-commerce main images and scene images, unification of brand visual assets, upgrading existing low-quality product photos, and design workflows that require multiple rounds of retouching while keeping the subject and layout stable. The Sunburst variant trades a longer single-run time for detail and lighting fidelity, making it suitable for output tasks where precision comes first.