Local image generation on Mac: 10 models compared (SD 1.5 → Flux dev → Qwen-Image → Gemini)

Tested 10 image generation models on M1 Max 64GB for photorealism, text rendering, and cultural accuracy (Japanese/Asian content).

Key findings:

Qwen-Image Lightning (8-step distillation) beats the full model in quality while being 9x faster (10min vs 93min)
Flux dev is the best local model for photorealism, but has strong English-centric bias (puts cilantro in ramen, turns izakayas into teahouses)
Gemini nails kanji rendering and cultural context, but it's cloud
SDXL Turbo generates in 5 seconds but quality is rough

The cultural accuracy gap surprised me most. Training data geography matters way more than model size for non-English content.