import torch
from diffusers import DiffusionPipeline
from peft import PeftModel

pipe = DiffusionPipeline.from_pretrained(
    "Tongyi-MAI/Z-Image-Turbo",
    dtype=torch.bfloat16,
    device_map="cuda"
)

pipe.load_lora_weights("SkyAsl/Pixel-artist-Z")
pipe.fuse_lora(lora_scale=0.85)

prompt = "a pixel art character with square orange glasses, a chef hat-shaped head and a purple-colored body on a cool background"

image = pipe(
    prompt=prompt,
    num_inference_steps=4,
    guidance_scale=0.0
).images[0]

image.save("pixel_art.png")

🧪 Evaluation

CLIPScore (Prompt–Image Alignment)

Model	Normalized CLIPScore (mean ± std)
Z-Image-Turbo	7.834 ± 2.577
Pixel-artist-Z	8.856 ± 2.473

+1.02 CLIPScore improvement

📈 Training Details

Dataset: m1guelpf/nouns
Image resolution: 512×512
Epochs: 1
Optimizer: AdamW (lr=1e-4)
Precision: bfloat16
Noise scheduler: DDPM (300 steps)

Slope: -1.1016515720251527e-05

LoRA Configuration

text

r = 16
lora_alpha = 32
lora_dropout = 0.05
target_modules = [to_q, to_k, to_v, to_out.0]

🖼️ Example Prompts

text

a pixel art character with a wizard hat and glowing blue eyes
a pixel art character holding a sword and wearing red armor
a pixel art character with a robot body and green visor

⚠️ Limitations

Optimized primarily for pixel art characters
May not improve (or may slightly degrade) photorealistic prompts