Owmen/Smallwater

This is a codebook tokenizer trained on Francesco/underwater-objects-5v7p8 dataset with 12k 16x16 tiles specifically for water. Below is an example of a reconstruction (Image is from the underwater-objects dataset)

Heres the code I used to reconstruct images:

import numpy as np import cv2 import faiss

codebook = np.load("codebook_12k.npy").astype('float32') index = faiss.IndexFlatL2(8) index.add(codebook)

def reconstruct_image(image_path, output_path="reconstructed.png"): img = cv2.imread(image_path) img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) img = cv2.resize(img, (640, 640))

patches = img.reshape(40, 16, 40, 16, 3).swapaxes(1, 2).reshape(-1, 16, 16, 3)

patches_q = (patches / 255.0 * 9).astype(np.float32)
mean_c = patches_q.mean(axis=(1, 2))
std_c = patches_q.std(axis=(1, 2))
bright = patches_q.mean(axis=(1, 2, 3))
sat = np.std(mean_c, axis=1)
feats = np.column_stack([mean_c, std_c, bright, sat])

_, token_ids = index.search(feats, k=1)
token_ids = token_ids.flatten()

reconstructed_patches = []
for tid in token_ids:
    mean_rgb = codebook[tid][:3] * (255.0 / 9.0)

    block = np.full((16, 16, 3), mean_rgb, dtype=np.uint8)
    reconstructed_patches.append(block)

recon_img = np.array(reconstructed_patches).reshape(40, 40, 16, 16, 3)
recon_img = recon_img.swapaxes(1, 2).reshape(640, 640, 3)

cv2.imwrite(output_path, cv2.cvtColor(recon_img, cv2.COLOR_RGB2BGR))
print(f"Test complete! Check {output_path}")

reconstruct_image("your_image.jpg")