Qwen Image 2512 Turbo LoRA

Qwen-Image-2512-Turbo-LoRA is a 4 or 8-steps turbo LoRA for Qwen Image 2512 trained by Wuli Team. This LoRA matches the original model's ouput quality but is over 20x faster⚡️, 2x from CFG-distillation and others from reduced number of inference steps.

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

You can also try this model on this DEMO webiste: https://modelscope.cn/studios/kelseye/Qwen-Image-2512-Turbo-LoRA-Demo

News

• Jan 8, 2026: Rlease V3.0 with further improved image color and details, please check the differences between V1.0, V2.0, and V3.0 in the first table below. This version is already ComfyUI compatible.
• Jan 2, 2026: Release V2.0 with improved image color and details. This version is already ComfyUI compatible.
• Dec 31, 2025: Release V1.0 and its ComfyUI compatible version (thanks to @System36), we will soon release V2.0 with enhanced performance.

Result Visualization

Differences between different versions

Prompt	Our LoRA V1.0 (4 steps)	Our LoRA V2.0 (4 steps)	Our LoRA V3.0 (4 steps)
ultra-realistic 3D render of four mechanical keyboard keycaps in a tight 2x2 grid, all keys touching. View from an isometric angle. One key is transparent with the word “Qwen” printed in white key. The other three colors are: black, purple, and white. The black key says the white “Image” . The other two say “25” and “12”. Realistic plastic texture, rounded sculpted keycaps, soft shadows, clean light-gray background.
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.
A dreamy and ethereal hand-drawn flat illustration in a Post-Impressionist style, featuring impressionistic brushwork and abstract, minimalist lines. A close-up view shows a little boy in plush pajamas balancing on a ladder made of clouds in the night sky. He is hanging freshly washed, wet stars that are dripping liquid light, one by one, onto a long clothesline strung between the tips of a crescent moon. Beside him, a glowing little rabbit is helping by handing him clothespins. The scene is filled with bright, vibrant colors, bokeh brushstrokes, washes of pale golden mist, soft textures, and gentle soft lighting with a soft focus effect.
Bookstore window display. A sign displays “New Arrivals This Week”. Below, a shelf tag with the text “Best-Selling Novels Here”. To the side, a colorful poster advertises “Author Meet And Greet on Saturday” with a central portrait of the author. There are four books on the bookshelf, namely “The light between worlds” “When stars are scattered” “The slient patient” “The night circus”
A four-panel sci-fi comedy comic strip, vertical layout. The style mixes futuristic cyberpunk elements with a mundane kitchen setting. Bright neon accents. Panel 1 (Top): A sleek, advanced humanoid robot with glowing blue eyes stands in a normal kitchen, wearing a "KISS THE COOK" apron. It holds a spatula dramatically. Text bubble (Robot, robotic font): "任务已接受：正在执行‘制作煎蛋’程序。成功率计算中：99.9%。" (Task Accepted: Executing 'Make Omelet' protocol. Calculating success rate: 99.9%.) Panel 2: The robot is staring intensely at a carton of eggs. Its eyes are projecting complex holographic scanning grids and analytical data over a single egg. Text bubble (Robot thinking): "分析蛋壳结构……探测微小裂缝……优化敲击力度矢量。" (Analyzing shell structure... detecting micro-fractures... optimizing impact force vectors.) Panel 3: CHAOS. The robot uses way too much force or advanced weaponry. It is firing a miniature laser beam from its finger at the egg, which has exploded into a cloud of shell and yolk. The kitchen is covered in mess. Text bubble (Sound effect, huge): "轰！！" (BOOM!!) Text bubble (Robot): "哎呀。" (Oops.) Panel 4 (Bottom): The robot stands covered in egg yolk, looking dejected. On the plate is a tiny, charred, unrecognizable black crisp. Text bubble (Robot): "任务失败。重新计算成功率：0.01%。我需要下载‘常识’补丁。" (Task Failed. Recalculating success rate: 0.01%. I need to download the 'Common Sense' patch.)

Differences between our turbo LoRA and the original model

Prompt	Qwen-Image-2512 (40 steps)	Qwen-Image-2512 + Our LoRA V1.0 (4 steps)	Qwen-Image-2512 + Our LoRA V1.0 (8 steps)
ultra-realistic 3D render of four mechanical keyboard keycaps in a tight 2x2 grid, all keys touching. View from an isometric angle. One key is transparent with the word “Qwen” printed in white key. The other three colors are: black, purple, and white. The black key says the white “Image” . The other two say “25” and “12”. Realistic plastic texture, rounded sculpted keycaps, soft shadows, clean light-gray background.
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.
A dreamy and ethereal hand-drawn flat illustration in a Post-Impressionist style, featuring impressionistic brushwork and abstract, minimalist lines. A close-up view shows a little boy in plush pajamas balancing on a ladder made of clouds in the night sky. He is hanging freshly washed, wet stars that are dripping liquid light, one by one, onto a long clothesline strung between the tips of a crescent moon. Beside him, a glowing little rabbit is helping by handing him clothespins. The scene is filled with bright, vibrant colors, bokeh brushstrokes, washes of pale golden mist, soft textures, and gentle soft lighting with a soft focus effect.
Bookstore window display. A sign displays “New Arrivals This Week”. Below, a shelf tag with the text “Best-Selling Novels Here”. To the side, a colorful poster advertises “Author Meet And Greet on Saturday” with a central portrait of the author. There are four books on the bookshelf, namely “The light between worlds” “When stars are scattered” “The slient patient” “The night circus”
A four-panel sci-fi comedy comic strip, vertical layout. The style mixes futuristic cyberpunk elements with a mundane kitchen setting. Bright neon accents. Panel 1 (Top): A sleek, advanced humanoid robot with glowing blue eyes stands in a normal kitchen, wearing a "KISS THE COOK" apron. It holds a spatula dramatically. Text bubble (Robot, robotic font): "任务已接受：正在执行‘制作煎蛋’程序。成功率计算中：99.9%。" (Task Accepted: Executing 'Make Omelet' protocol. Calculating success rate: 99.9%.) Panel 2: The robot is staring intensely at a carton of eggs. Its eyes are projecting complex holographic scanning grids and analytical data over a single egg. Text bubble (Robot thinking): "分析蛋壳结构……探测微小裂缝……优化敲击力度矢量。" (Analyzing shell structure... detecting micro-fractures... optimizing impact force vectors.) Panel 3: CHAOS. The robot uses way too much force or advanced weaponry. It is firing a miniature laser beam from its finger at the egg, which has exploded into a cloud of shell and yolk. The kitchen is covered in mess. Text bubble (Sound effect, huge): "轰！！" (BOOM!!) Text bubble (Robot): "哎呀。" (Oops.) Panel 4 (Bottom): The robot stands covered in egg yolk, looking dejected. On the plate is a tiny, charred, unrecognizable black crisp. Text bubble (Robot): "任务失败。重新计算成功率：0.01%。我需要下载‘常识’补丁。" (Task Failed. Recalculating success rate: 0.01%. I need to download the 'Common Sense' patch.)

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", path="Wuli-Qwen-Image-2512-Turbo-LoRA-4steps-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": None
}
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=4,  # 8 is also recommended
    seed=42,
    width=1328,
    height=1328
)
output.save("output.png")

Existing problem

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