Qwen-Image-2512-Turbo-LoRA-2-Steps

<div align="center"> <img src="https://img.alicdn.com/imgextra/i4/O1CN011xTzi6280yXyGmk9b_!!6000000007871-0-tps-3072-938.jpg" width=300 /> </div>

This is a 2-step turbo LoRA for Qwen Image 2512 trained by Wuli Team, representing an advancement over our 4-step turbo LoRA.

For users in Chinese mainland, you can try our website: https://wuli.art/generate, getting four images with 2k resolution generated by Qwen Image 2512 Turbo with only 5 seconds.

<div align="center"> <img src="https://img.alicdn.com/imgextra/i4/O1CN01rE2RnD1DmfhZD9e54_!!6000000000259-2-tps-1600-2859.png" width=500 /> </div>

Result Visualization

Quick start with Diffsynth-Engine

• Diffsynth-Engine：https://github.com/modelscope/DiffSynth-Engine
• Example Code:

import math
from diffsynth_engine import fetch_model, QwenImagePipeline, QwenImagePipelineConfig

# Create pipeline
config = QwenImagePipelineConfig.basic_config(
    model_path=fetch_model("Qwen/Qwen-Image-2512", path="transformer/*.safetensors"),
    encoder_path=fetch_model("Qwen/Qwen-Image-2512", path="text_encoder/*.safetensors"),
    vae_path=fetch_model("Qwen/Qwen-Image-2512", path="vae/*.safetensors"),
    offload_mode="cpu_offload",
)
pipe = QwenImagePipeline.from_pretrained(config)

# Load our turbo LoRA
pipe.load_lora(
    path=fetch_model("Wuli-art/Qwen-Image-2512-Turbo-LoRA-2-Steps", path="Wuli-Qwen-Image-2512-Turbo-LoRA-2steps-V1.0-bf16.safetensors"),
    scale=1.0,
    fused=True,
)

# Change scheduler config
scheduler_config = {
    "exponential_shift_mu": math.log(2.5),
    "use_dynamic_shifting": True,
    "shift_terminal": 0.7155
}
pipe.apply_scheduler_config(scheduler_config)

# Sample image
output = pipe(
    prompt="a young girl with flowing long hair, wearing a white halter dress and smiling sweetly. The background features a blue seaside where seagulls fly freely.",
    cfg_scale=1,
    num_inference_steps=2,
    seed=42,
    width=1328,
    height=1328
)
output.save("output.png")

Existing problem

• Generated image quality may degrade when handling extremely complex text rendering tasks, you can increase the value of num_inference_steps.