app.py

import pickle
import gradio as gr
from datasets import load_dataset
from transformers import AutoModel, AutoFeatureExtractor
from PIL import Image

# Only runs once when the script is first run.
with open("butts_1024_new.pickle", "rb") as handle:
    index = pickle.load(handle)

# Load model for computing embeddings.
feature_extractor = AutoFeatureExtractor.from_pretrained(
    "sasha/autotrain-butterfly-similarity-2490576840"
)
model = AutoModel.from_pretrained("sasha/autotrain-butterfly-similarity-2490576840")

# Candidate images.
dataset = load_dataset("sasha/butterflies_10k_names_multiple")
ds = dataset["train"]


def query(image, top_k=1):
    inputs = feature_extractor(image, return_tensors="pt")
    model_output = model(**inputs)
    embedding = model_output.pooler_output.detach()
    results = index.query(embedding, k=top_k)
    inx = results[0][0].tolist()
    logits = results[1][0].tolist()
    butterfly = ds.select(inx)["image"][0]
    return overlay_png_on_side_by_side_images(
        image, butterfly, "cadre_hq.png", png_position=(0, 0)
    )


def overlay_png_on_side_by_side_images(
    person_image_path, insect_image_path, overlay_image_path, png_position=(0, 0)
):
    """
    Overlays a PNG image onto two side-by-side images, resizing them to fit.

    Args:
        image1_path (str): Path to the first image.
        image2_path (str): Path to the second image.
        png_path (str): Path to the PNG image to overlay.
        output_path (str): Path to save the resulting image.
        png_position (tuple): (x, y) coordinates for the top-left corner of the PNG overlay.
    """
    # Open images
    img1 = Image.fromarray(person_image_path).convert("RGBA")
    img2 = insect_image_path.convert("RGBA")
    png_img = Image.open(overlay_image_path).convert("RGBA")
    min_height = 3600
    new_width = 2700
    img1_hpercent = min_height / float(img1.size[1])
    img1_wsize = int((float(img1.size[0]) * float(img1_hpercent)))
    img2_hpercent = min_height / float(img2.size[1])
    img2_wsize = int((float(img2.size[0]) * float(img2_hpercent)))

    img1 = img1.resize((img1_wsize, min_height), Image.LANCZOS)
    im1_left = (img1_wsize - new_width) / 2
    im1_right = (img1_wsize + new_width) / 2
    top = 0
    bottom = min_height
    cropped_img1 = img1.crop((int(im1_left), int(top), int(im1_right), int(bottom)))

    img2 = img2.resize((img2_wsize, min_height), Image.LANCZOS)
    im2_left = (img2_wsize - new_width) / 2
    im2_right = (img2_wsize + new_width) / 2
    cropped_img2 = img2.crop((int(im2_left), int(top), int(im2_right), int(bottom)))

    combined_width = 5400
    # Create a blank canvas for the combined image

    combined_image = Image.new(
        "RGBA", (combined_width, min_height), (0, 0, 0, 0)
    )  # Transparent background

    # Paste images side by side
    combined_image.paste(cropped_img1, (0, 0))
    combined_image.paste(cropped_img2, (cropped_img1.width, 0))

    # Resize PNG to fit within the combined image dimensions if necessary, or to a desired size
    # For simplicity, let's resize the PNG to a quarter of the combined image's width, maintaining aspect ratio
    target_png_width = combined_image.width
    png_img = png_img.resize((combined_image.width, min_height), Image.LANCZOS)

    # Overlay the PNG image
    combined_image.paste(png_img, png_position, png_img)
    return combined_image


with gr.Blocks() as demo:
    gr.Markdown("# Find my Butterfly 🦋 Trouver mon papillon 🦋")
    with gr.Row():
        with gr.Column(scale=1):
            inputs = gr.Image(width=288, height=384)
            btn = gr.Button("🦋")
            description = gr.Markdown()

        with gr.Column(scale=2):
            outputs = gr.Image(format="jpg")

    gr.Markdown("### Image Examples")
    gr.Examples(
        examples=["elton.jpg", "ken.jpg", "gaga.jpg", "taylor.jpg"],
        inputs=inputs,
        outputs=outputs,
        fn=query,
        cache_examples=True,
    )
    btn.click(query, inputs, outputs)

demo.launch()