app.py

import gradio as gr

from transformers import AutoModel

import torch
import numpy as np
from PIL import Image

MAX_IMAGES = 4


def generate_small(color_indexed: bool, color_num: int) -> list:
    """Generates a small sprite.

    Parameters
    ----------
    color_indexed : bool
        Whether to use color indexing.
    color_num : int
        Number of colors in the palette.

    Returns
    -------
    list
        List of PIL images.
    """
    # Get the latent dimension
    latent_dim = model_small.model.latent_dim
    # Initialize the list of images
    images_list = []
    # Generate MAX_IMAGES images
    for _ in range(MAX_IMAGES):
        # Generate a random latent vector
        latents = torch.randn((1, latent_dim))
        # Generate the image
        with torch.no_grad():
            generated_image = model_small(latents)
        # Clamp the image to [0, 1]
        generated_image = generated_image.clamp_(0.0, 1.0).cpu().numpy()

        # Convert the generated image to PIL image
        color_image = Image.fromarray(
            np.uint8(generated_image[0] * 255).transpose(1, 2, 0), "RGB"
        )

        # Convert to color indexed image if needed
        if color_indexed:
            # Convert using adaptive palette of given color depth
            color_image_indexed = color_image.convert(
                "P", palette=Image.ADAPTIVE, colors=color_num
            )
            # Add the color indexed image to the list
            images_list.append(color_image_indexed)

        # Add the image to the list
        images_list.append(color_image)

    return images_list


def generate_med(color_indexed: bool, color_num: int) -> list:
    """Generates a medium sprite.

    Parameters
    ----------
    color_indexed : bool
        Whether to use color indexing.
    color_num : int
        Number of colors in the palette.

    Returns
    -------
    list
        List of PIL images.
    """
    # Get the latent dimension
    latent_dim = model_med.model.latent_dim
    # Initialize the list of images
    images_list = []
    # Generate MAX_IMAGES images
    for _ in range(MAX_IMAGES):
        # Generate a random latent vector
        latents = torch.randn((1, latent_dim))
        # Generate the image
        with torch.no_grad():
            generated_image = model_med(latents)
        # Clamp the image to [0, 1]
        generated_image = generated_image.clamp_(0.0, 1.0).cpu().numpy()

        # Convert the generated image to PIL image
        color_image = Image.fromarray(
            np.uint8(generated_image[0] * 255).transpose(1, 2, 0), "RGBA"
        )

        # Convert to color indexed image if needed
        if color_indexed:
            # Convert using adaptive palette of given color depth
            color_image_indexed = color_image.convert(
                "P", palette=Image.ADAPTIVE, colors=color_num
            )
            # Add the color indexed image to the list
            images_list.append(color_image_indexed)

        # Add the image to the list
        images_list.append(color_image)

    return images_list


# Create the demo interface
demo = gr.Blocks()

# Create the small model
model_small = AutoModel.from_pretrained(
    "michaelriedl/MonsterForge-small", trust_remote_code=True
)
model_small.eval()

# Create the medium model
model_med = AutoModel.from_pretrained(
    "michaelriedl/MonsterForge-medium", trust_remote_code=True
)
model_med.eval()

# Create the interface
with demo:
    gr.HTML(
        """
        <div style="text-align: center; margin: 0 auto;">
            <p style="margin-bottom: 14px; line-height: 23px;">
                Gradio demo for MonsterForge models. This was built with Lightweight GAN using the implementation from <a href='https://github.com/lucidrains/lightweight-gan' target='_blank'>lucidrains</a>.
            </p>
        </div>
        """
    )
    with gr.Tabs():
        with gr.TabItem("Small Sprite"):
            with gr.Column():
                with gr.Row():
                    gallery_small = gr.Gallery(
                        columns=4,
                        object_fit="scale-down",
                    )
                with gr.Row():
                    color_index_small = gr.Checkbox(label="Color indexed", value=False)
                    color_num_small = gr.Slider(
                        minimum=8,
                        maximum=32,
                        value=32,
                        step=4,
                        label="Number of colors in the palette",
                    )
                gen_btn_small = gr.Button("Generate")
                gen_btn_small.click(
                    fn=generate_small,
                    inputs=[color_index_small, color_num_small],
                    outputs=gallery_small,
                )
        with gr.TabItem("Medium Sprite"):
            with gr.Column():
                with gr.Row():
                    gallery_med = gr.Gallery(
                        columns=4,
                        object_fit="scale-down",
                    )
                with gr.Row():
                    color_index_med = gr.Checkbox(label="Color indexed", value=False)
                    color_num_med = gr.Slider(
                        minimum=8,
                        maximum=32,
                        value=32,
                        step=4,
                        label="Number of colors in the palette",
                    )
                gen_btn_med = gr.Button("Generate")
                gen_btn_med.click(
                    fn=generate_med,
                    inputs=[color_index_med, color_num_med],
                    outputs=gallery_med,
                )
    gr.HTML(
        """
        <div class="footer">
            <div style='text-align: center;'>MonsterForge by <a href='https://michaelriedl.com/' target='_blank'>Michael Riedl</a></div>
        </div>
        """
    )

# Launch the interface
demo.launch()