Mac image benchmark crowns Qwen Lightning

This reads less like a model-size contest and more like a reminder that training data geography matters more than raw scale when prompts leave the English default.

• –Qwen-Image Full looks strong on kanji, but 93 minutes per image makes it impractical; the 8-step Lightning distillation is the real winner because it is much faster and sometimes better composed.
• –Flux dev remains the best local pick for photorealism, but its Western bias shows up in obvious prompt drift on food and cultural scenes.
• –Gemini is still the cleanest option for text and cultural accuracy, but the cloud dependency makes it a different deployment class.
• –SDXL Turbo is only useful when latency matters more than fidelity; it is a prototyping model, not a deliverable generator.