Qwen

/

Qwen-Image-Edit-2509

like 1.15k

Follow

Qwen 86.6k

Here is the complete code from community of Qwen-Image_edit.I modified it slightly

import torch
from transformers import BitsAndBytesConfig as TransformersBitsAndBytesConfig
from transformers import Qwen2_5_VLForConditionalGeneration

from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig
from diffusers import QwenImageEditPipeline, QwenImageTransformer2DModel,QwenImageEditPlusPipeline
from diffusers.utils import load_image


model_id = "Qwen/Qwen-Image-Edit-2509"
torch_dtype = torch.bfloat16
device = "cuda"

quantization_config = DiffusersBitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    llm_int8_skip_modules=["transformer_blocks.0.img_mod"],
)

transformer = QwenImageTransformer2DModel.from_pretrained(
    model_id,
    subfolder="transformer",
    quantization_config=quantization_config,
    torch_dtype=torch_dtype,
)
transformer = transformer.to("cpu")

quantization_config = TransformersBitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)


text_encoder = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_id,
    subfolder="text_encoder",
    quantization_config=quantization_config,
    torch_dtype=torch_dtype,
)
text_encoder = text_encoder.to("cpu")

pipe = QwenImageEditPlusPipeline.from_pretrained(
    model_id, transformer=transformer, text_encoder=text_encoder, torch_dtype=torch_dtype
)

#optionally load LoRA weights to speed up inference
pipe.load_lora_weights("lightx2v/Qwen-Image-Lightning", weight_name="Qwen-Image-Lightning-8steps-V1.1.safetensors")
pipe.load_lora_weights(
    "lightx2v/Qwen-Image-Lightning", weight_name="Qwen-Image-Lightning-4steps-V1.0-bf16.safetensors"
)
pipe.enable_model_cpu_offload()

generator = torch.Generator(device="cuda").manual_seed(21)

image = load_image(
    "/path/to/your/image.jpg"
).convert("RGB")


prompt = "your prompt here"
# change steps to 8 or 4 if you used the lighting loras
image = pipe(image, prompt, num_inference_steps=8).images[0]

image.save("qwenimageedit.png")