(3) The same message applies to several models: Chroma, Z image, Klein, Ernie, Midjourney

An independent AI researcher has conducted a viral, methodical comparison testing seven leading open-source and proprietary image generation models against the gold-standard aesthetic of Midjourney. The models under scrutiny included multiple versions of Chroma (V41, V48, Radiance, Alpha), Z Image Turbo, Klein 9b Turbo, and Ernie Turbo. The core of the test was a single, highly detailed cinematic prompt describing a lone traveler ascending ancient stairs toward a divine, swirling vortex of golden light. Crucially, the prompt was first rewritten by a large language model (LLM) to inject a level of creative styling akin to Midjourney's own prompt interpretation, rather than using the raw text across all models. This approach aimed to level the playing field for aesthetic output.

The researcher's primary metric was how closely each model could aesthetically replicate the distinctive, beautiful, and detailed style that has made Midjourney a benchmark. Tests for models like Z Image Turbo and Klein 9b included runs both with and without LoRAs (Low-Rank Adaptations), which are small add-on files that modify a model's output style. The researcher explicitly excluded quantized versions of models (like Qwen) to ensure a fair comparison of each model's full, uncompromised quality. The results, shared in a detailed Reddit post, provide a rare side-by-side look at the current capabilities of these models in a high-stakes fantasy realism scenario, offering valuable insights for developers and artists choosing between rapidly evolving AI image tools.