app.py

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

# Load the pre-trained diffusion model
base_model = "ptx0/pixart-900m-1024-ft-v0.7-stage1"
stg2_model = "ptx0/pixart-900m-1024-ft-v0.7-stage2"
torch_device = "cuda"
base_pipeline = PixArtSigmaPipeline.from_pretrained(
	base_model, use_safetensors=True
).to(dtype=torch_precision, device=torch_device)
stg2_pipeline = PixArtSigmaPipeline.from_pretrained(stg2_model, **base_pipeline.components)
stg2_pipeline.transformer = PixArtTransformer2DModel.from_pretrained(stg2_model, subfolder="transformer").to(dtype=torch_precision, device=torch_device)
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)
    mixture_generator = torch.Generator().manual_seed(444)
    stage1_strength = 0.6
    latent_images = base_pipeline(
        prompt=prompt,
        negative_prompt=negative_prompt,
        num_inference_steps=num_inference_steps,
        num_images_per_prompt=1,
        generator=mixture_generator,
        guidance_scale=guidance_scale,
        output_type="latent",
        denoising_end=stage1_strength,
        width=width,
        height=height
    ).images
    return stg2_pipeline(
        prompt=prompt,
        negative_prompt=negative_prompt,
        latents=latent_images,
        num_inference_steps=num_inference_steps,
        num_images_per_prompt=1,
        generator=mixture_generator,
        guidance_scale=guidance_scale,
        denoising_start=stage1_strength
    ).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 two-stage mixture-of-experts model implemented in the spirit of NVIDIA's E-Diffi model."
        "<br />The weights were initialised from <strong>terminusresearch/pixart-900m-1024-ft-v0.6</strong> and trained separately on timestep ranges <strong>999-400</strong> and <strong>400-0</strong>."
        "<br />This results in two models where the first stage is responsible for most of the image's composition and colour, and the second stage handles minor-to-fine details."
        "<br />"
        "<br />In comparison to SDXL's refiner, the second stage here handles twice as many timesteps, which allows it to make more use of the text-conditional guidance, improving its capabilities."
        "<br />"
        "<br />Despite being trained with 40% of the schedule, you will discover that using stage 2 stand-alone as a refiner (img2img) will need half the strength - about 20%."
        "<br />When being used in the two-stage pipeline, it should be configured to handle all of its 40% range."
        "<br />"
        "<br />This model is funded and trained by <strong>Terminus Research Group</strong>."
        " If you would like to collaborate or provide compute, please see the organisation page for how to locate us on Discord."
        "<br />"
        "<br />"
        "<ul>"
        "<li>Lead trainer: @pseudoterminalx (bghira@GitHub)</li>"
        "</ul>"
    )
).launch()