Open image generation models are closer to closed-source quality than this sub thinks [D]

I run evaluations on generative image models as part of my workflow, mostly comparing coherence, prompt adherence, and compositional accuracy across different architectures. The consensus here seems to be that open models are still a generation behind closed APIs. Based on my recent benchmarks, that gap is way smaller than people assume. On compositional control specifically, the latest open checkpoints handle multi-object scenes with spatial relationships about as reliably as the paid endpoints I've tested