app.py

import gradio as gr
import spaces
import torch
from diffusers import Transformer2DModel
from scripts.diffusers_patches import pixart_sigma_init_patched_inputs, PixArtSigmaPipeline

assert getattr(Transformer2DModel, '_init_patched_inputs', False), "Need to Upgrade diffusers: pip install git+https://github.com/huggingface/diffusers"
setattr(Transformer2DModel, '_init_patched_inputs', pixart_sigma_init_patched_inputs)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
weight_dtype = torch.float16

transformer = Transformer2DModel.from_pretrained(
    "PixArt-alpha/PixArt-Sigma-XL-2-1024-MS",
    subfolder='transformer',
    torch_dtype=weight_dtype,
    use_safetensors=True,
)
pipe = PixArtSigmaPipeline.from_pretrained(
    "PixArt-alpha/pixart_sigma_sdxlvae_T5_diffusers",
    transformer=transformer,
    torch_dtype=weight_dtype,
    use_safetensors=True,
)
pipe.to(device)

@spaces.GPU(duration=90)
def generate(prompt, negative_prompt, num_inference_steps, guidance_scale, height, width):
    image = pipe(
        prompt,
        negative_prompt=negative_prompt,
        num_inference_steps=num_inference_steps,
        guidance_scale=guidance_scale,
        height=height,
        width=width
    ).images[0]
    return image

interface = gr.Interface(
    fn=generate,
    inputs=[
        gr.Text(label="Prompt"),
        gr.Text(label="Negative Prompt"),
        gr.Slider(minimum=10, maximum=100, value=30, step=1, label="Number of Inference Steps"),
        gr.Slider(minimum=1, maximum=20, value=6, step=0.1, label="Guidance Scale"),
        gr.Slider(minimum=64, maximum=1024, value=1024, step=64, label="Height"),
        gr.Slider(minimum=64, maximum=1024, value=1024, step=64, label="Width"),
    ],
    outputs=gr.Image(label="Generated Image"),
    title="PixArt Sigma Image Generation",
    description="""Generate high-fidelity 4K images from text prompts using PixArt-Sigma, a state-of-the-art diffusion model.

PixArt-Sigma achieves superior image quality and alignment with prompts compared to previous models like [PixArt-alpha](https://github.com/PixArt-alpha/PixArt-sigma). It does so efficiently, evolving from PixArt-alpha through a process termed weak-to-strong training - leveraging higher quality data and an improved attention mechanism.

With just 0.6 billion parameters, PixArt-Sigma reaches new heights in text-to-image generation. Output stunning, intricate 4K images for posters, wallpapers, concept art, and more. Guide the model with descriptive prompts and fine-tune parameters like guidance scale and number of inference steps.

For more information, visit the original [repository](https://github.com/PixArt-alpha/PixArt-sigma) and follow the HF Space creator on Twitter at [@artificialguybr](https://twitter.com/artificialguybr).
""",
)

interface.launch()