samoudi/sd-turbo-gguf

SD-Turbo — GGUF (q8_0 / q4_0)

Powered by Stability AI.

GGUF conversions of stabilityai/sd-turbo for stable-diffusion.cpp, used by the offline Android app Offline Image Gen (on-device text-to-image, CPU only, no network after the first download).

Files

File	Quantization	Size (bytes)	SHA-256	Recommended RAM
sd_turbo_q8_0.gguf	q8_0	2149136352	604cfbd79e72706884027ea9e883c2daf9b243598b71daa19b15871f6d8dc15f	≥ 6 GB
sd_turbo_q4_0.gguf	q4_0	1847909248	d069ef2b2d3d5c2465662d706c242a1016db56163c109bd7e3008a94cdd00359	≥ 4 GB

The UNet convolutions are not quantized by stable-diffusion.cpp and stay in f16, which is why q4_0 is only ~14 % smaller than q8_0.

Source and conversion

• Source: stabilityai/sd-turbo, file sd_turbo.safetensors (SHA-256 3f067a1b943cf162f2b8f8588f6cf5824bd5b4c7d1d88d87164b9ca123616549).
• Tool: stable-diffusion.cpp sd-cli, tag master-955-b7a74e6 (official Windows CPU build).
• Commands (two passes, so that non-quantized tensors are stored in f16 instead of f32):

sd-cli -M convert -m sd_turbo.safetensors -o sd_turbo-f16.gguf --type f16
sd-cli -M convert -m sd_turbo-f16.gguf   -o sd_turbo_q8_0.gguf --type q8_0
sd-cli -M convert -m sd_turbo-f16.gguf   -o sd_turbo_q4_0.gguf --type q4_0

Usage

SD-Turbo is distilled for 1–4 steps without guidance:

sd-cli -m sd_turbo_q8_0.gguf -p "a white house with a red roof, garden, blue sky, photo" \
  --steps 4 --cfg-scale 1.0 --sampling-method euler_a -W 512 -H 512 --seed 42 -o out.png

Prompts should be written in English. 512×512 is the native resolution; 256×256 is faster.

In the Offline Image Gen app

On first launch, the app downloads one of these files (chosen from the device RAM) through the direct link https://huggingface.co/samoudi/sd-turbo-gguf/resolve/main/<file>, with resume support (HTTP Range), checks its size and SHA-256, then works fully offline.

License

Distributed under the Stability AI Community License (copy in LICENSE.md, attribution in NOTICE.txt). Free for research, non-commercial and commercial use for individuals and organizations under USD 1M in annual revenue; above that, an enterprise license from Stability AI is required. Use must comply with the Stability AI Acceptable Use Policy.