Definitely. I run an entire site built around a series of benchmarks that focus on prompts of increasingly difficult complexity with a focus on adherence, and even the state-of-the-art local models are probably only about thirty percent as good as proprietary models like Gemini 3.1 Flash Image and GPT Image 2.

Comparing Qwen-Image, Flux.2, ZiT, NB2, and gpt-image-2

https://genai-showdown.specr.net/?models=qi,nbp3,f2d,g2,zt