alexturkey/z-image-turbo-pascal-4gb
Media & DesignRun Z-Image-Turbo Q8_0 text-to-image at 768x1344 on a 4 GB Pascal laptop GPU under Windows.
What it is
This project packages a working recipe for running Z-Image-Turbo (Q8_0 GGUF) text-to-image generation at 768x1344 on a 4 GB GTX 1050 Ti laptop card (Pascal, sm_61) under Windows, using stable-diffusion.cpp with a cuDNN-accelerated VAE stage. The reported test machine is an i7-7700HQ laptop with 16 GB RAM, Windows 10, driver 581.29, models on SSD. It is unofficial and not affiliated with Tongyi-MAI, leejet or NVIDIA.
The value lies less in the model than in the engine patches: a refiner block that ran twice per step is fixed, the VAE decode moves to cuDNN float32, and decoding tries a single piece first, then two large tiles, then small tiles. The README reports 193–228 s per image and a VAE stage of 8.3–15.9 s in v0.2.0, versus 261.6 s and 29.8–36.3 s in the unpublished v0.1.0 layout (comparison table).
Quick start
Requirements per the README: Windows 10/11, an NVIDIA 4 GB GPU, a CUDA 12.8-capable driver (581.29 measured), Python 3.10+, git, ~14 GB free disk and 16 GB RAM (requirements).
git clone https://github.com/alexturkey/z-image-turbo-pascal-4gb.git
cd z-image-turbo-pascal-4gb
python scripts/setup.py
python scripts/generate.py --prompt-file prompts/eiffel.txt
setup.py prints a Preflight block, downloads models from pinned addresses in models.lock.json and verifies them with sha256, printing MODEL_OK, ENGINE_OK, CUDART_OK, then CUDNN_OK or CUDNN_SKIPPED, and finally SETUP_OK. Downloads total about 12.2 GB with cuDNN. cuDNN is NVIDIA software installed from NVIDIA's server by default; --no-cudnn skips it and falls back to the slower --vae legacy path. Interrupted downloads resume from .part files.
generate.py writes outputs/<timestamp>_<name>/image.png, plus run.json (timings, VAE path, peak GPU memory, exact command) and log.txt, and prints IMAGE_OK <path>. Other options include --prompt, --width/--height (multiples of 16, 256–2048), --seed, --out, --name, --models-dir, --bin-dir and --vae auto|tiled|legacy (usage).
What it runs
The diffusion model is z_image_turbo-Q8_0.gguf from leejet's GGUF conversion of the base Tongyi-MAI model; the text encoder is Qwen3-4B-Instruct-2507 Q8_0; the VAE is ae.safetensors from Comfy-Org. The engine is a patched stable-diffusion.cpp build (32d516c), shipped prebuilt as sd-cli.exe in the repository's v0.2.0 release, with a CUDA 12.8 runtime from the wan22-pascal-4gb project and cuDNN 9.10.2.21 downloaded separately (the parts table). Settings are fixed: cfg 1.0, 8 steps, euler sampler, simple scheduler, --offload-to-cpu --mmap --eager-load --flash-attn on.
Constraints
- Windows only, with a prebuilt engine compiled only for Pascal sm_61. Other GPU architectures need your own build via BUILD.md; that path is untested by the author.
- Only the GTX 1050 Ti 4 GB laptop card and driver 581.29 were measured. Other GPUs, Linux, macOS, and machines with under 16 GB RAM are untested; community notes for some other cards are in OTHER_GPUS.md.
- Memory is tight: the process holds roughly 10.7–10.9 GB of RAM, leaving a few hundred MiB free; run one image at a time.
- Only 768x1344, 1344x768 and 1024x1024 were tested; larger sizes warn and may run out of memory.
- cuDNN and the CUDA runtime are NVIDIA's property under NVIDIA's license, not this repository's Apache-2.0. Model files are downloaded from their upstream sources under their own licenses and are not stored in the repo. Generated images have no stated ownership terms (license section).
- Prompt fidelity is described as partial; text rendering and real landmarks were not evaluated.
HotRepo's take
On the HotRepo new-project list for 16 November 2026, the repository reached 106 stars with 6 gained inside a roughly 1.5-hour measured window — a small, early sample rather than a trend. What stands out in the material is the measurement discipline: each claimed number is tied to a command and a file, the pixel hash lets readers check reproducibility, and the limits section names what was not tested (other cards, drivers, less RAM, non-Pascal builds). The gap worth knowing about is that the speed and VAE improvements are the author's patches, while fitting into 4 GB comes from stable-diffusion.cpp itself, and the unpatched upstream engine was not run on this card. For readers with exactly this hardware class, the recipe is concrete; for everyone else it is a starting point for a build of their own, not a drop-in package.
Written by AI from the project's README, updated . The project's own docs are the source of truth.