app.py

import os
import gradio as gr
import torch
import numpy as np
import random
from diffusers import StableDiffusion3Pipeline, AutoencoderKL, SD3Transformer2DModel, FlowMatchEulerDiscreteScheduler
import spaces
from PIL import Image
import requests
import transformers
from translatepy import Translator

os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
translator = Translator()
HF_TOKEN = os.environ.get("HF_TOKEN", None)
# Constants
model = "stabilityai/stable-diffusion-3-medium"
repo= "stabilityai/stable-diffusion-3-medium-diffusers"
MAX_SEED = np.iinfo(np.int32).max

CSS = """
.gradio-container {
  max-width: 690px !important;
}
footer {
    visibility: hidden;
}
"""

JS = """function () {
  gradioURL = window.location.href
  if (!gradioURL.endsWith('?__theme=dark')) {
    window.location.replace(gradioURL + '?__theme=dark');
  }
}"""


vae = AutoencoderKL.from_pretrained(
    repo,
    subfolder="vae",
    torch_dtype=torch.float16,
)

transformer = SD3Transformer2DModel.from_pretrained(
    repo, 
    subfolder="transformer",
    torch_dtype=torch.float16,
)

text_encoder_3 = T5EncoderModel.from_pretrained(
    repo, 
    subfolder="text_encoder_3", 
    torch_dtype=torch.float16,
)

# Ensure model and scheduler are initialized in GPU-enabled function
if torch.cuda.is_available():
    pipe = StableDiffusion3Pipeline.from_pretrained(repo, vae=vae, transformer=transformer, text_encoder_3=text_encoder_3, torch_dtype=torch.float16).to("cuda")

pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_config(pipe.scheduler.config)

# Function 
@spaces.GPU()
def generate_image(
    prompt,
    negative="low quality",
    width=1024,
    height=1024,
    scale=1.5,
    steps=28,
    clip=3):

    
    seed = random.randint(0, MAX_SEED)
    generator = torch.Generator().manual_seed(seed)
    
    prompt = str(translator.translate(prompt, 'English'))

    print(f'prompt:{prompt}')
    
    image = pipe(
        prompt, 
        negative_prompt=negative, 
        width=width,
        height=height,
        guidance_scale=scale,
        num_inference_steps=steps,
        clip_skip=clip,
        generator = generator,
    )
    return image.images[0]


examples = [
    "a cat eating a piece of cheese",
    "a ROBOT riding a BLUE horse on Mars, photorealistic",
    "Ironman VS Hulk, ultrarealistic",
    "a CUTE robot artist painting on an easel",
    "Astronaut in a jungle, cold color palette, oil pastel, detailed, 8k",
    "An alien holding sign board contain word 'Flash', futuristic, neonpunk",
    "Kids going to school, Anime style"
]


# Gradio Interface

with gr.Blocks(css=CSS, js=JS, theme="soft") as demo:
    gr.HTML("<h1><center>SD3M🦄</center></h1>")
    gr.HTML("<p><center><a href='https://huggingface.co/stabilityai/stable-diffusion-3-medium'>sd3m</a> text-to-image generation</center><br><center>Multi-Languages. Adding default prompts to enhance.</center></p>")
    with gr.Group():
        with gr.Row():
            prompt = gr.Textbox(label='Enter Your Prompt', value="best quality, HD, aesthetic", scale=6)
            submit = gr.Button(scale=1, variant='primary')
    img = gr.Image(label='SD3M Generated Image')
    with gr.Accordion("Advanced Options", open=False):
        with gr.Row():
            negative = gr.Textbox(label="Negative prompt", value="low quality")
        with gr.Row():
            width = gr.Slider(
                label="Width",
                minimum=512,
                maximum=1280,
                step=8,
                value=1024,
            )
            height = gr.Slider(
                label="Height",
                minimum=512,
                maximum=1280,
                step=8,
                value=1024,
            )
        with gr.Row():
            scale = gr.Slider(
                label="Guidance",
                minimum=3.5,
                maximum=7,
                step=0.1,
                value=5,
            )
            steps = gr.Slider(
                label="Steps",
                minimum=1,
                maximum=50,
                step=1,
                value=28,
            )
            clip = gr.Slider(
                label="Clip Skip",
                minimum=1,
                maximum=10,
                step=1,
                value=3,
            )      
    gr.Examples(
        examples=examples,
        inputs=prompt,
        outputs=img,
        fn=generate_image,
        cache_examples="lazy",
    )

    prompt.submit(fn=generate_image,
                 inputs=[prompt, negative, width, height, scale, steps, clip],
                 outputs=img,
                 )
    submit.click(fn=generate_image,
                 inputs=[prompt, negative, width, height, scale, steps, clip],
                 outputs=img,
                 )
    
demo.queue().launch()