Open image generation models are closer to closed-source quality than this sub thinks [D]
An individual running evaluations on generative image models suggests that open-source models are rapidly closing the quality gap with closed-source alternatives. Based on benchmarks for compositional accuracy and prompt adherence, the latest open checkpoints perform comparably to paid APIs. Furthermore, text rendering in generated images has improved significantly, with open models achieving correct rendering 70-80% of the time for short strings. AI
IMPACT Suggests open-source image generation models are becoming competitive with proprietary options, potentially lowering barriers for developers.