|
|
|
|
|
import torch |
|
|
from diffusers import StableDiffusion3Pipeline |
|
|
import random |
|
|
|
|
|
class ImageGenerator: |
|
|
def __init__(self, repo: str): |
|
|
self.repo = repo |
|
|
self.pipeline = StableDiffusion3Pipeline.from_pretrained(self.repo, torch_dtype=torch.float16, local_files_only = True) |
|
|
self.pipeline.enable_model_cpu_offload() |
|
|
|
|
|
def generate_image(self, prompt: str, width: int = 1024, height: int = 1024, scale_factor: float = 4.5, steps: int = 28, seed: int = None): |
|
|
seed = seed if seed is not None else random.randint(0, 2**32 - 1) |
|
|
print(f"using {seed} to generate image...") |
|
|
generator = torch.Generator("cuda").manual_seed(seed) |
|
|
image = self.pipeline( |
|
|
prompt, |
|
|
negative_prompt="", |
|
|
width=width, |
|
|
height=height, |
|
|
num_inference_steps=steps, |
|
|
guidance_scale=scale_factor, |
|
|
max_sequence_length=512, |
|
|
generator=generator, |
|
|
).images[0] |
|
|
return image |
|
|
|
|
|
Sd_repo = "/tmp/pretrainmodel/stable-diffusion-3.5-medium-ungated" |
|
|
sd_model = ImageGenerator(Sd_repo) |
|
|
prompt = "A close-up portrait of an Asian girl with blunt bangs and big eyes, side profile, holding a red apple on top of her head, in a winter beach setting. She looks very happy, with snowflakes gently falling on her hair. The scene is captured with a high-quality DSLR camera, showcasing natural light and bokeh effects, with a fresh, crisp light and shadow play, reminiscent of a snowy film scene." |
|
|
image = sd_model.generate_image(prompt) |
|
|
|
