Z-Image Turbo for Intel Vulkan
A smaller Windows Vulkan option for text-to-image. This recipe pins the Q3_K denoiser, a Q3_K_S Qwen3 4B encoder, and the full VAE. It downloads 5.37 GB of model assets and requests a 6 GiB runtime budget. The weights come from their original publishers and are checked against pinned SHA-256 digests.
Test results
On a Windows 11 Intel Arc B390 with driver 32.0.101.9033 and stable-diffusion.cpp 532d2c1, the first 512ร512 image arrived in 20.01 seconds from server launch. The second request took 12.92 seconds. The runtime assigned 4,952 MB of parameters to Vulkan and none to host memory. This figure excludes execution buffers and driver overhead.
The fox was coherent. The poster captured the fern and general layout, but misspelled and rearranged some of the requested lettering. For typography, the MOSS HaDeS recipe did better on the same prompt. Iris Xe timing and driver behavior still need a hardware pass.
Run on Windows
Use 64-bit Python 3.10 or newer and an Intel graphics driver with Vulkan.
py -m pip install huggingface_hub
hf download xgirls/Z-Image-Turbo-Intel-Vulkan run.py presets.json --local-dir fern-z-image
cd fern-z-image
py run.py --benchmark
py run.py --prompt "A fox beside a woodland stream" --output my-image
The benchmark generates the two included samples at 512ร512, eight steps, CFG 1, and fixed seeds. It saves PNGs, timings, and a Vulkan log. Use a new --output directory per run.
Sources
Z-Image Turbo, denoiser GGUF by leejet, Qwen3 4B encoder GGUF by Unsloth, and VAE by Comfy-Org.
Model tree for xgirls/Z-Image-Turbo-Intel-Vulkan
Base model
Tongyi-MAI/Z-Image-Turbo