Layer Extraction (Qwen-Image-2.1 LoRA)

This model is designed to extract specific elements from an image. The extracted layers are saved as images with transparent backgrounds, making them well-suited for detailed editing workflows.

• Training Framework: DiffSynth-Studio
• Base Model: Qwen-Image-2.1
• Dataset: PrismLayersPro
• Training Resources: 8 ZW810 PPUs
• Training Steps: 60,000 steps

Suggested prompt template: Extract the following object: xxx

Showcase

Input Image	Prompt	Output Image
	Extract the following object: A girl with wings.
	Extract the following object: Wings.
	Extract the following object: A girl holding two boxes.
	Extract the following object: A yellow hat.

Inference Code

Install DiffSynth-Studio:

git clone https://github.com/modelscope/DiffSynth-Studio.git
cd DiffSynth-Studio
pip install -e .

Load the models and run inference:

from diffsynth.pipelines.qwen_image_21 import QwenImage21Pipeline, ModelConfig
import torch
from PIL import Image
from modelscope import snapshot_download

vram_config = {
    "offload_dtype": "disk",
    "offload_device": "disk",
    "onload_dtype": "disk",
    "onload_device": "disk",
    "preparing_dtype": torch.bfloat16,
    "preparing_device": "cuda",
    "computation_dtype": torch.bfloat16,
    "computation_device": "cuda",
}
pipe = QwenImage21Pipeline.from_pretrained(
    torch_dtype=torch.bfloat16,
    device="cuda",
    model_configs=[
        ModelConfig(model_id="Qwen/Qwen-Image-2.1", origin_file_pattern="transformer/diffusion_pytorch_model*.safetensors", **vram_config),
        ModelConfig(model_id="Qwen/Qwen-Image-2.1", origin_file_pattern="text_encoder/model*.safetensors", **vram_config),
        ModelConfig(model_id="Qwen/Qwen-Image-2.1", origin_file_pattern="vae/diffusion_pytorch_model*.safetensors", **vram_config),
    ],
    processor_config=ModelConfig(model_id="Qwen/Qwen-Image-2.1", origin_file_pattern="processor/"),
    vram_limit=torch.cuda.mem_get_info("cuda")[1] / (1024 ** 3) - 0.5,
)
pipe.enable_lora_hot_loading(pipe.dit)

# Extract layers using prompt
snapshot_download("DiffSynth-Studio/Qwen-Image-2.1-LayerExtract", allow_file_pattern="assets/*", local_dir="data")
pipe.load_lora(pipe.dit, ModelConfig(model_id="DiffSynth-Studio/Qwen-Image-2.1-LayerExtract", origin_file_pattern="model.safetensors"))

prompt = "Extract the following object: A girl with wings."
image = pipe(prompt, seed=0, height=1024, width=1024,
             edit_image=Image.open("data/assets/image_input_2.png"))
image.save("image_extract_girl_wings.png")

pipe.clear_lora()