source/app.py

from __future__ import annotations

from pathlib import Path

import gradio as gr
import numpy as np
import torch
from model import ConditionalVAE
from PIL import Image
from safetensors.torch import load_file

ARTIFACT = Path(__file__).resolve().parent / "artifacts" / "glyph-forge-cvae"
MODEL = ConditionalVAE()
MODEL.load_state_dict(load_file(ARTIFACT / "model.safetensors"))
MODEL.eval()


def generate_digit(label: int, seed: int) -> Image.Image:
    generator = torch.Generator().manual_seed(seed)
    latent = torch.randn(1, MODEL.latent_dimensions, generator=generator)
    with torch.no_grad():
        pixels = MODEL.decode(latent, torch.tensor([label]))[0]
    image = np.clip(pixels.reshape(8, 8).numpy() * 255, 0, 255).astype(np.uint8)
    return Image.fromarray(image, mode="L").resize((512, 512), Image.Resampling.NEAREST)


with gr.Blocks(title="GlyphForge CVAE") as demo:
    gr.Markdown("# GlyphForge\nSample the eight-dimensional latent style of any digit.")
    with gr.Row():
        label = gr.Slider(0, 9, value=3, step=1, label="Digit")
        seed = gr.Slider(0, 100_000, value=2031, step=1, label="Latent seed")
    output = gr.Image(value=generate_digit(3, 2031), label="Generated 8x8 glyph")
    button = gr.Button("Forge glyph", variant="primary")
    button.click(generate_digit, inputs=[label, seed], outputs=output)


if __name__ == "__main__":
    demo.launch()