app.py

# %%

import cv2
from sklearn.cluster import KMeans
from PIL import Image
import numpy as np
import gradio.components as gc
import gradio as gr

 
def pixart(
    i,
    block_size=4,
    n_clusters=5,
    hsv_weights=[0, 0, 1],
    local_contrast_blur_radius=51,  # has to be odd
    upscale=True,
    seed=None,
    output_scaling=1,
    dither_amount=15
):
    w, h = i.size
    dw = w//block_size
    dh = h//block_size

    # always resize with NEAREST to keep the original colors
    i = i.resize((dw, dh), Image.Resampling.NEAREST)
    ai = np.array(i)

    if seed is None:
        # seed = np.random.randint(0, 2**32 - 1)
        seed = np.random.randint(0, 2**16 - 1)
    km = KMeans(n_clusters=n_clusters, random_state=seed)

    hsv = cv2.cvtColor(ai, cv2.COLOR_RGB2HSV)
    bhsv = cv2.GaussianBlur(
        hsv,
        (local_contrast_blur_radius, local_contrast_blur_radius),
        0,
        borderType=cv2.BORDER_REPLICATE
    )
    hsv32 = hsv.astype(np.float32)
    km.fit(
        hsv32.reshape(-1, hsv32.shape[-1]),
        # (sharp-blurred) gives large values if a pixel stands out from its surroundings
        # raise to the power of 4 to make the difference more pronounced.
        # this preserves rare specks of color by increasing the probability of them getting their own cluster
        sample_weight=(
            np.linalg.norm((hsv32 - bhsv), axis=-1).reshape(-1)
            ** 4
        )
    )
    label_grid = km.labels_.reshape(hsv32.shape[:2])
    centers = km.cluster_centers_ # hsv values

    def pick_representative_pixel(cluster):
        '''pick the representative pixel for a cluster'''
        most_sat_color = (hsv[label_grid == cluster] @
                          np.array(hsv_weights)).argmax()
        return hsv[label_grid == cluster][most_sat_color]
    
    cluster_colors = np.array([
        pick_representative_pixel(c)
        for c in range(centers.shape[0])])
    


    if dither_amount == 0:
        # assign each pixel the color of its cluster
        ki = cluster_colors[label_grid]
    else:
        # add noise to the colors before selecting the nearest color, this acts as a dithering effect
        noised_colors = hsv32 + np.random.normal(0, dither_amount, hsv.shape)
        noised_colors = np.clip(noised_colors, 0, 255)
        flattened = noised_colors.reshape(-1, 3)
        # use the dot product to find the closest cluster (could also try euclidean distance)
        closest_clusters = np.argmax(flattened @ centers.T,axis=1)
        closest_clusters_eucledian = np.argmin(np.linalg.norm(centers - flattened[:, None], axis=-1), axis=1)
        label_grid = closest_clusters_eucledian.reshape(hsv32.shape[:2])
        ki = cluster_colors[label_grid]

    rgb = cv2.cvtColor(ki.astype(np.uint8), cv2.COLOR_HSV2RGB)
    i = Image.fromarray(rgb)
    if upscale:
        i = i.resize((w, h), Image.Resampling.NEAREST)

    if output_scaling != 1:
        i = i.resize(
            (w*output_scaling, h*output_scaling), Image.Resampling.NEAREST)
    return i, seed


def query(
    i: Image.Image,
    block_size: str,
    n_clusters,  # =5,
    hsv_weights,  # ='0,0,1'
    local_contrast_blur_radius,  # =51 has to be odd
    seed,  # =42,
    output_scaling,
    dither_amount
):
    bs = float(block_size)
    w, h = i.size
    if bs < 1:
        blsz = int(bs * min(w, h))
    else:
        blsz = int(bs)

    hw = [float(w) for w in hsv_weights.split(',')]

    pxart, usedseed = pixart(
        i,
        block_size=blsz,
        n_clusters=n_clusters,
        hsv_weights=hw,
        local_contrast_blur_radius=local_contrast_blur_radius,
        upscale=True,
        seed=int(seed) if seed != '' else None,
        output_scaling=output_scaling,
        dither_amount=dither_amount
    )
    if n_clusters <= 256:
        pxart = pxart.convert('P', palette=Image.Palette.ADAPTIVE, colors=n_clusters)
    #pxart.save('temp.bmp')
    return pxart, usedseed


# %%
searchimage = gc.Image(
    # shape=(512, 512),
    label="Search image", type='pil')
block_size = gc.Textbox(
    "0.01",
    label='Block Size ',
    placeholder="e.g. 8 for 8 pixels. 0.01 for 1% of min(w,h) (<1 for percentages, >= 1 for pixels)")
palette_size = gc.Slider(
    1, 1024, 32, step=1, label='Palette Size (Number of Colors)')
hsv_weights = gc.Textbox(
    "0,0,1",
    label='HSV Weights. Weights of the channels when selecting a "representative pixel"/centroid from a cluster of pixels',
    placeholder='e.g. 0,0,1 to only consider the V channel (which seems to work well)')
lcbr = gc.Slider(
    3, 512, 51, step=2, label='Blur radius to calculate local contrast')

seed = gc.Textbox(
    "",
    label='Seed for the random number generator (empty to randomize)',
    placeholder='e.g. 42')

outimage = gc.Image(
    # shape=(224, 224), 
    label="Output", type='pil')
seedout = gc.Textbox(label='used seed')

output_scaling = gc.Slider(
    0, 16, 1, step=1, label='Output scaling factor')

dither_amount = gc.Slider(
    0, 255, 0, step=1, label='Dithering amount')

gr.Interface(
    query,
    [searchimage, block_size, palette_size, hsv_weights,  lcbr, seed, output_scaling, dither_amount],
    [outimage, seedout],
    title="kmeans-Pixartifier",
    description=f"Turns images into pixel art using kmeans clustering",
    analytics_enabled=False,
    allow_flagging='never',
).launch()