app.py

import spaces
import torch
from diffusers import PixArtSigmaPipeline
import gradio as gr

# Load the pre-trained diffusion model
pipe = PixArtSigmaPipeline.from_pretrained('ptx0/pixart-900m-1024-ft', torch_dtype=torch.bfloat16)
pipe.to('cuda')
import re

def extract_resolution(resolution_str):
    match = re.match(r'(\d+)x(\d+)', resolution_str)
    if match:
        width = int(match.group(1))
        height = int(match.group(2))
        return (width, height)
    else:
        return None

# Define the image generation function with adjustable parameters and a progress bar
@spaces.GPU
def generate(prompt, guidance_scale, num_inference_steps, resolution, negative_prompt):
    width, height = extract_resolution(resolution) or (1024, 1024)
    return pipe(
        prompt,
        negative_prompt=negative_prompt,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        width=width, height=height
    ).images

# Example prompts to demonstrate the model's capabilities
example_prompts = [
    ["A futuristic cityscape at night under a starry sky", 7.5, 25, "blurry, overexposed"],
    ["A serene landscape with a flowing river and autumn trees", 8.0, 20, "crowded, noisy"],
    ["An abstract painting of joy and energy in bright colors", 9.0, 30, "dark, dull"]
]

# Create a Gradio interface, 1024x1024,1152x960,896x1152
iface = gr.Interface(
    fn=generate,
    inputs=[
        gr.Text(label="Enter your prompt"),
        gr.Slider(1, 20, step=0.1, label="Guidance Scale", value=3.4),
        gr.Slider(1, 50, step=1, label="Number of Inference Steps", value=28),
        gr.Radio(["1024x1024", "1152x960", "896x1152"], label="Resolution", value="1152x960"),
        gr.Text(value="underexposed, blurry, ugly, washed-out", label="Negative Prompt")
    ],
    outputs=gr.Gallery(height=1024, min_width=1024, columns=2),
    examples=example_prompts,
    title="PixArt 900M",
    description=(
        "This is a 900M parameter model expanded from PixArt Sigma 1024px (600M) by adding 14 layers to deepen the transformer."
        "<br />This model is being <strong>actively trained</strong> on 3.5M samples across a wide distribution of photos, synthetic data, cinema, anime, and safe-for-work furry art."
        "<br />"
        "<br />&nbsp;The datasets been filtered for extremist and illegal content, but it is possible to produce toxic outputs. <strong>This model has not been safety-aligned or fine-tuned</strong>."
        " You may receive non-aesthetic results, or prompts might be partially or wholly ignored."
        "<br />Although celebrity names and artist styles haven't been scrubbed from the datasets, the low volume of these samples in the training set result in a lack of representation for public figures."
        "<br />"
        "<br />Be mindful when using this demo space that you do not inadvertently share images without adequate preparation and informing the receivers that these images are AI generated."
        "<br />"
        "<br />This model was developed and funded by <strong>Terminus Research Group</strong>."
        "<br />"
        "<br />"
        "<ul>"
        "<li>Lead trainer: @pseudoterminalx (bghira@GitHub)</li>"
        "<li>Architecture: @jimmycarter (AmericanPresidentJimmyCarter@GitHub)</li>"
        "<li>Datasets: @ProGamerGov, @jimmycarter, @pseudoterminalx</li>"
        "</ul>"
    )
).launch()