mks0813/z-image-turbo-pixel-art-lora

Z-Image-Turbo Pixel Art LoRA

A LoRA adapter for Tongyi-MAI/Z-Image-Turbo fine-tuned on pixel art style images.

Model Details

• Base Model: Tongyi-MAI/Z-Image-Turbo
• Training Framework: DiffSynth Studio
• LoRA Rank: 32
• Target Modules: to_q, to_k, to_v, to_out.0, w1, w2, w3
• Training Dataset: mks0813/pixel_image_dataset

Usage

Installation

pip install diffsynth

Inference Code

Important: Add , pxlstl suffix to your prompt for best pixel art style results.

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

MODEL_ID = "Tongyi-MAI/Z-Image-Turbo"
STYLE_SUFFIX = ", pxlstl"

# Load pipeline
pipe = ZImagePipeline.from_pretrained(
    torch_dtype=torch.bfloat16,
    device="cuda",
    model_configs=[
        ModelConfig(model_id=MODEL_ID, origin_file_pattern="transformer/*.safetensors"),
        ModelConfig(model_id=MODEL_ID, origin_file_pattern="text_encoder/*.safetensors"),
        ModelConfig(model_id=MODEL_ID, origin_file_pattern="vae/diffusion_pytorch_model.safetensors"),
    ],
    tokenizer_config=ModelConfig(model_id=MODEL_ID, origin_file_pattern="tokenizer/"),
)

# Load LoRA
from huggingface_hub import hf_hub_download
lora_path = hf_hub_download(repo_id="mks0813/z-image-turbo-pixel-art-lora", filename="z-image-turbo-pixel-art-lora.safetensors")
pipe.load_lora(pipe.dit, lora_path)

# Generate image
prompt = "A cute pixel art cat" + STYLE_SUFFIX
image = pipe(prompt=prompt, seed=42, rand_device="cuda")
image.save("output.png")

Sample Outputs

Image	Prompt
	A cat is driving a car in the city
	A cozy café on a rainy Paris street at night
	A dancer spinning on a rooftop under city lights
	A fox sleeping in a sunlit autumn forest
	A iconic white-sand beach lined with high-rise hotels and swaying palm trees
	A landscape of a meadow with sheep wandering at sunset
	A lighthouse standing against crashing waves under a stormy sky
	A lion is sleeping under the tree
	A lone astronaut walking on a crystal desert planet

Training Details

• Epochs: 2
• Learning Rate: 1e-4
• Gradient Checkpointing: Enabled
• Dataset Size: 492 images

License

This model follows the license of the base model Tongyi-MAI/Z-Image-Turbo.