Comparison of low Steps, Klein 9b x Z image turbo x Ernie Turbo x Qwen 2512 8 Steps

An independent AI art enthusiast conducted a real-world performance test of four leading open-source image generation models, all capable of running on consumer hardware with 8GB of VRAM. The models—Klein 9B (6 steps), Z Image Turbo (9 steps), Ernie Turbo (8 steps), and Qwen 2512 (8 steps, Q4KM quantized)—were benchmarked on generation speed, with all completing renders in under 40 seconds. The test used a consistent 832x1216 resolution and simple prompts without style-specific LoRAs to create a direct, 'brutal' comparison of their base capabilities. The goal was to evaluate performance in a typical user scenario, prioritizing speed and quality for a practical workflow.

Results revealed clear strengths for each model. Klein 9B and Z Image Turbo maintained their dominance in generating realistic human subjects, with Klein noted as particularly effective for adult-themed content using LoRAs. The newcomer, Baidu's Ernie Turbo, was the surprise standout, producing an aesthetic in some images that closely resembled the high-quality output of Midjourney. Notably, Ernie Turbo also operated with less built-in censorship than Klein 9B. While Qwen 2512 offered impressive style diversity, its output was marked as more recognizably 'AI-like,' a flaw the tester attributed to its aggressive quantization. The test highlights the rapid evolution of efficient, locally-runnable models that challenge the quality of closed, subscription-based services.