wbruna/TwinFlow-Z-Image-Turbo-sdcpp-GGUF

Quantized GGUF versions of the TwinFlow Z-Image Turbo model, for stable-diffusion.cpp.

Converted from https://huggingface.co/azazeal2/TwinFlow-Z-Image-Turbo-repacked with stable-diffusion.cpp. Example conversion:

./sd-cli --mode convert --model TwinFlow_Z_Image_Turbo_exp_bf16.safetensors --tensor-type-rules "^context_refiner.*(attention\.(out|qkv)|feed_forward).*weight=q8_0,^(layers|noise_refiner).*(adaLN_modulation|attention\.(out|qkv)|feed_forward).*weight=q5_0"  --output TwinFlow_Z_Image_Turbo_exp-Q5_0.gguf

(note, I didn't test these with ComfyUI, it may or may not work!)

Model Information

See the original model card at https://huggingface.co/inclusionAI/TwinFlow-Z-Image-Turbo

Usage

You need at least release master-385-34a6fd4. Get also the VAE and LLM model from https://huggingface.co/leejet/Z-Image-Turbo-GGUF .

Example command:

./sd-cli --diffusion-model TwinFlow_Z_Image_Turbo_exp-Q4_0.gguf --vae ae_bf16.safetensors --llm qwen_3_4b-Q8_0.gguf --cfg-scale 1 --steps 3 -p "an apple"

Parameters:

• cfg-scale 1
• sampling_method euler (default), with scheduler discrete (default), smoothstep or sgm_uniform, and 2-4 steps
• sampling_method dpm2, with scheduler smoothstep or sgm_uniform, and 2-3 steps

For low VRAM setups, you may follow How to Use Z‐Image on a GPU with Only 4GB VRAM, changing the number of steps to 2-5.

Credits

• TwinFlow Z-Image Turbo
• Z-Image Turbo by Tongyi-MAI
• Safetensors model: https://huggingface.co/azazeal2/TwinFlow-Z-Image-Turbo-repacked
• GGUF files quantized with mainline stable-diffusion.cpp

License

Apache 2.0