kpsss34/Qwen-Image-8PR

kpsss34/Qwen-Image-8PR is a model that cuts out unimportant parts (in the text_encoder part), thus making the model smaller in size.

from diffusers import QwenImagePipeline
import torch

model_name = "kpsss34/Qwen-Image-8PR"

if torch.cuda.is_available():
    torch_dtype = torch.bfloat16
    device = "cuda"
else:
    torch_dtype = torch.float32
    device = "cpu"

pipe = QwenImagePipeline.from_pretrained(model_name, torch_dtype=torch_dtype)

pipe = pipe.to(device)
#pipe.enable_model_cpu_offload() #If you have less than 40GB of vram

positive_magic = {
    "en": ", Ultra HD, 4K, cinematic composition.",
    "zh": ", 超清，4K，电影级构图."
}

prompt = '''a medium shot of a woman with long blonde hair, wearing a green sweater and looking to the left. her eyes are slightly open, and she has a serious expression on her face. her lips are red, and her teeth are slightly visible'''

negative_prompt = " "

aspect_ratios = {
    "1:1": (1024, 1024),
    "16:9": (1280, 720),
    "9:16": (720, 1280),
}

width, height = aspect_ratios["1:1"]

image = pipe(
    prompt=prompt + positive_magic["en"],
    negative_prompt=negative_prompt,
    width=width,
    height=height,
    num_inference_steps=30,
    true_cfg_scale=4.0,
    generator=torch.Generator(device="cuda").manual_seed(42)
).images[0]

image.save("example.png")