Krea-2-Turbo • GGUF & Humming quantizations

Image Comparison Explorer

Twenty-four prompts from the Qwen-Image-Bench 2K set, each rendered by every model of both families at 8 steps, cfg 1.0, euler sampling, flow shift 1.15, and a fixed seed. Pick any two variants to compare side by side or with a slider, and zoom in to inspect text, faces, and fine texture.

Krea-2-Turbo (GGUF) Krea-2-Turbo (Humming)

Back to the ByteShape blog

All prompts

GGUF Humming

Prompt 1 of 24

Prompt Show full

Side by side Slider

100% Reset

A

B

A

B

Scroll or double‑click to zoom • drag to pan

Choose two variants Swap A / B

Prompts come from the Qwen-Image-Bench prompt set (Qwen/Qwen-Image-Bench), 2K split. Every image was generated at 8 steps, cfg 1.0, euler, flow shift 1.15, with a fixed seed per prompt and the BF16 Qwen3-VL-4B text encoder. GGUF images come from the ComfyUI / stable-diffusion.cpp pipeline, Humming images from vLLM-Omni; each family is compared against its own BF16 baseline. Full-resolution images are downscaled to a 1792 px long edge for the web. Treat these images as a fair comparison across variants, not as the best each model can do.