Generative AI Image-Guided Editing Benchmarks (springus.io)

A new benchmark evaluates image-guided generative-editing models on one-shot transfers that must combine a “fit pic” (the subject/garment) with a “query” flat‑lay pose while preserving identity-defining details—text, logos, printed graphics and complex patterns—and using a templated text prompt. Tasks span graphic reconstruction, pattern reconstruction, small-segment enhancement and multi-image fusion; all models were tested under controlled conditions (same prompt, best-of-3 generations, 1024×1024 JPEG, square crops, identical test images). Nano Banana Pro led with 8/12 passing cases (Nano Banana 7/12), followed by GPT Image‑1 (4), Seedream 4 (3) and Qwen (2). The benchmark biases toward consistency over isolated best outputs to stress true one-shot performance. Technically, the Nano Banana family excels at preserving logos, patterns and pose transfer in single-shot edits, showing strong color fidelity and clean background transfer; failure modes include minor edge sharpening, placement/scale errors and multi-angle text reconstruction problems. GPT Image‑1 showed strength in multi-image alignment but weaker graphic/text fidelity, suggesting it may benefit from few-shot inputs. Key implications: current top models are production‑viable for one-shot e‑commerce edits but trade quality for cost/latency (Nano Banana Pro’s advantage likely offset by higher inference cost), and robust few‑shot/multi‑image handling remains the biggest unmet need. Future work should expand few‑shot and multi‑image tests to better reflect real app scenarios.