Generative AI Image Editing Showdown (genai-showdown.specr.net)

A public “generative AI image editing showdown” put six leading multimodal editors through a battery of realistic, targeted tasks — from “give this bald man a full thick head of hair” and swapping blocks’ positions to complex multi-change edits (turn JAWS into PAWS) and preserving painterly styles like The Great Wave and Girl with a Pearl Earring. Results were mixed: Gemini 2.5 Flash often excelled at element placement (surfer, card color change) but sometimes only changed colors or globally altered images; OpenAI’s gpt-image-1 frequently rewrote entire scenes and is constrained by stricter I/O resolutions; FLUX.1 Kontext dev surprisingly outperformed its larger counterpart on several tasks but showed instability (many retries/looping); OmniGen2 and Qwen-Image-Edit made plausible edits yet tended to introduce collateral facial or compositional changes; Seedream 4 occasionally succeeded where others failed (shortening the giraffe) but often warped style or texture (e.g., “Stalin” hair, weathered sign retention). The contest highlights core technical pain points: regional, instruction-faithful edits remain hard without unintended global changes; iterative edits compound degradations; maintaining fine-grained texture, lighting, and spatial constraints is inconsistent; and model behavior varies widely with architecture, training data, and resolution handling. Practical implications: current editors are useful for bold compositing but risky for subtle restorations or high-assurance workflows. The results argue for better benchmarking, support for multi-image reference inputs, improved region-aware conditioning, and stronger constraints to reduce collateral edits.