baptle/Z-Image-Japan-Expert

Japan Realistic LoRA for Z-Image-Turbo

A LoRA adapter trained on realistic Japanese photography to enhance Z-Image-Turbo's ability to generate authentic Japanese scenes, urban landscapes, and cultural elements.

Model Description

This is a LoRA (Low-Rank Adaptation) adapter trained on the Tongyi-MAI/Z-Image-Turbo diffusion model. It specializes in generating realistic photographs of Japanese locations, transportation, architecture, and everyday scenes with authentic lighting and composition.

Training Details

• Base Model: Tongyi-MAI/Z-Image-Turbo
• Training Steps: 2,000
• LoRA Rank (r): 32
• LoRA Alpha: 32
• Learning Rate: 0.0001
• Optimizer: AdamW 8-bit
• Batch Size: 1 (with gradient accumulation of 4)
• Training Resolution: 512x512
• Precision: bfloat16
• Noise Scheduler: FlowMatch
• Trained Using: Ostris AI-Toolkit

Usage

Using with Diffusers

from diffusers import DiffusionPipeline
import torch

# Load base model
pipe = DiffusionPipeline.from_pretrained(
    "Tongyi-MAI/Z-Image-Turbo",
    torch_dtype=torch.bfloat16
)
pipe.to("cuda")

# Load LoRA adapter
pipe.load_lora_weights("your-username/japan_realistic")

# Generate image
prompt = "Photo of a Shinkansen bullet train stopped at a Japanese station platform, overhead roof structure, yellow tactile paving, natural daylight, ultra realistic."
image = pipe(
    prompt=prompt,
    num_inference_steps=8,
    guidance_scale=1.0,
    width=1024,
    height=1024
).images

image.save("output.png")