app.py

PATH = 'harpomaxx/deeplili' #stable diffusion 1.5
from PIL import Image
import torch
import torch.distributed as dist
import torch.multiprocessing as mp
import argparse

#from diffusers import StableDiffusionPipeline, LMSDiscreteScheduler
from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler


from tqdm.auto import tqdm
import random
import gradio as gr

def generate_images(prompt, guidance_scale, n_samples, num_inference_steps):
    seeds = [random.randint(1, 10000) for _ in range(n_samples)]
    images = []
    for seed in tqdm(seeds):
        torch.manual_seed(seed)
        image = pipe(prompt, num_inference_steps=num_inference_steps,guidance_scale=guidance_scale).images[0]
        images.append(image)
    return images

def gr_generate_images(prompt: str, num_images = 1, num_inference = 20, guidance_scale = 8 ):
    prompt = prompt + "sks style"
    images = generate_images(prompt, guidance_scale, num_images, num_inference)
    return images

with gr.Blocks() as demo:
    examples = [
    [
        'A black and white cute character on top of a hill',
        1,
        30
    ],
    [
        'Bubbles and mountains in the sky',
        1,
        20
    ],
    [
        'A tree with multiple eyes and a small flower muted colors',
        1,
        20
    ],
    [
        "3d character on top of a hill",
        1,
        20
    ],
    [
        "a poster of a large forest with black and white characters",
        1,
        20
    ],
    ]
    gr.Markdown(
    """
    <img src="https://github.com/harpomaxx/DeepLili/raw/main/images/lilifiallo/660.png" width="150" height="150">

    # #DeepLili v0.5b

    ## Enter your prompt and generate a work of art in the style of Lili's Toy Art paintings.
    ## (English, Spanish)
    """
    )

    with gr.Column(variant="panel"):
        with gr.Row(variant="compact"):
            text = gr.Textbox(
                label="Enter your prompt",
                show_label=False,
                max_lines=2,
                placeholder="a white and black drawing of  a cute character on top of a house with a little animal"
            ).style(
                container=False,
            )

        with gr.Row(variant="compact"):
          #  num_images_slider = gr.Slider(
          #      minimum=1,
          #      maximum=10,
          #      step=1,
          #      value=1,
          #      label="Number of Images",
          #  )

          #  num_inference_steps_slider = gr.Slider(
          #      minimum=1,
          #      maximum=25,
          #      step=1,
          #      value=20,
          #      label="Inference Steps",
          #  )

          #  guidance_slider = gr.Slider(
          #      minimum=1,
          #      maximum=14,
          #      step=1,
          #      value=8,
          #      label="Guidance Scale",
          #  )



            btn = gr.Button("Generate image").style(full_width=False)

        #gallery = gr.Gallery(
        #    label="Generated images", show_label=False, elem_id="gallery"
        #).style(columns=[1], rows=[1], object_fit="scale-down", height="50%", width="50%")


        gallery = gr.Image(
            label="Generated image", 
            type='pil',
            show_label = False,
            tool=None,  # Disable editing tools if not needed
        ).style(
            width="50%",  # Use the full view width
            height="50%",  # Use the full view height
            object_fit="scale-down"  # Maintain aspect ratio without cropping
        )

    num_images_slider = 1
    num_inference_steps_slider = 20
    guidance_slider = 8

    btn.click(gr_generate_images, [text], gallery)
    gr.Examples(examples, inputs=[text])
    gr.HTML(
    """
    <h6><a href="https://harpomaxx.github.io/"> harpomaxx </a></h6>
    """
    )

if __name__ == "__main__":
    # Check if CUDA is available
    if torch.cuda.is_available():
        device = "cuda"
        dtype = torch.float16  # Use float16 for GPU to save memory
    else:
        device = "cpu"
        dtype = torch.float32  # CPU does not support float16, use float32 instead


    dpm = DPMSolverMultistepScheduler.from_pretrained(PATH, subfolder="scheduler")
    pipe = StableDiffusionPipeline.from_pretrained(PATH,torch_dtype=dtype, scheduler=dpm).to(device)
    
    demo.queue(concurrency_count=2,
            ).launch()