DiffSynth-Studio/Z-Image-Turbo-DistillPatch

Z-Image Turbo Acceleration Capability Fix LoRA

Model Introduction

This model is a LoRA used to fix the acceleration capability of Z-Image Turbo LoRA.

LoRAs trained directly based on Z-Image Turbo will lose their acceleration capability. Images generated under acceleration configuration (steps=8, cfg=1) become blurry, while images generated under non-acceleration configuration (steps=30, cfg=2) remain normal.

Results

Training Data:

Generation Results:

steps=8, cfg=1	steps=30, cfg=2	steps=8, cfg=1, with our model fix

Inference Code

from diffsynth.pipelines.z_image import ZImagePipeline, ModelConfig
import torch

pipe = ZImagePipeline.from_pretrained(
    torch_dtype=torch.bfloat16,
    device="cuda",
    model_configs=[
        ModelConfig(model_id="Tongyi-MAI/Z-Image-Turbo", origin_file_pattern="transformer/*.safetensors"),
        ModelConfig(model_id="Tongyi-MAI/Z-Image-Turbo", origin_file_pattern="text_encoder/*.safetensors"),
        ModelConfig(model_id="Tongyi-MAI/Z-Image-Turbo", origin_file_pattern="vae/diffusion_pytorch_model.safetensors"),
    ],
    tokenizer_config=ModelConfig(model_id="Tongyi-MAI/Z-Image-Turbo", origin_file_pattern="tokenizer/"),
)
pipe.load_lora(pipe.dit, "path/to/your/lora.safetensors")
pipe.load_lora(pipe.dit, ModelConfig(model_id="DiffSynth-Studio/Z-Image-Turbo-DistillPatch", origin_file_pattern="model.safetensors"))

image = pipe(prompt="a dog", seed=42, rand_device="cuda")
image.save("image.jpg")