product-photo-studio

Build a JSON manifest with one job per SKU, lock the prompt on 5 pilot SKUs, then run gen-ai batch run with a dry-run cost estimate first. Use recraftv3-replace-bg for pixel-accurate background replacement at scale.

Use qwen-image-edit-plus in image-to-image mode with the hero photo as reference and hex codes in the prompt, such as "same product recolored in midnight navy #1B2A44". One job per variant keeps lighting and composition identical.

recraftv3-replace-bg offers the best product fidelity for label text, logos, and fine detail. For regulated products like cosmetics or food, avoid flux-kontext-pro which can hallucinate small typography.

Yes, the mockups mode renders flat artwork onto merch like tees, mugs, and posters using gemini-3-pro-image with the artwork passed as an -i reference. Generate one surface per job and match aspect ratios to the platform.

Drift happens when prompts use vague terms like "warm" instead of exact hex codes and explicit shadow direction. Pin the exact color values, state "same lighting and shadow as reference", and regenerate outliers individually.

Avoid full reshoots for a handful of hero images, where direct gen-ai generate calls are simpler, and for color-critical products like jewelry close-ups where physical accuracy cannot be validated post-generation.