Qwen-Image-2512 • GGUF & Humming quantizations
Image Comparison Explorer
These examples compare and contrast images generated by the BF16 and quantized models, illustrating both successes and a range of failure modes. Each of the 24 prompts was expanded by the Qwen prompt refiner into a longer, more specific brief, and every image was sampled at 1024×1024 with 50 sampling steps, cfg 4, and one fixed seed per prompt shared across every variant. Two families are in the set: the GGUF builds run through ComfyUI / stable-diffusion.cpp, and the Humming builds run through vLLM-Omni with the Humming kernels.
Back to the Qwen-Image-2512 postAll 24 prompts
Thumbnails show the active family’s BF16 baseline.
What to look for
The 24 prompts are sampled from Qwen/Qwen-Image-Bench and expanded by the Qwen prompt refiner. Every image was generated at 1024×1024 with 50 steps, cfg 4, and one fixed seed per prompt shared across every variant; each family is compared against its own BF16 baseline. Treat these images as a fair comparison across variants, not as the best each model can do.