VQGAN-Encoded Image Features from CC12M Dataset

Approximately 1.5 million images from the Conceptual Captions 12M (CC12M) collection, encoded into feature representations using a VQGAN f16 1024 model. The dataset was created by johnowhitaker and last updated in April 2022. A script for further preprocessing is provided in the repository.

Use Cases

• Train generative models using the VQGAN-encoded image features as a compressed representation of the ~1.5M images.
• Fine-tune downstream vision models on tasks like classification or retrieval using the pre-extracted VQGAN features from the CC12M collection.
• Analyze the distribution of VQGAN feature codes across the image corpus to understand visual concept clustering.

Strengths

• Contains approximately 1.5 million encoded images, providing a substantial corpus for model training.
• Features are pre-extracted using a specific VQGAN f16 1024 model, offering a consistent and computationally efficient representation.
• Includes a script for further dataset preparation, offering flexibility for users.

Limitations

• The underlying image data is from 2022 or earlier, which may not reflect current visual trends or subjects.
• The dataset consists only of pre-computed encodings, not the original image files, limiting tasks that require pixel-level access.
• The specific quality, diversity, and potential biases of the original CC12M image collection are inherited by this encoded version.

Provenance

Source

Conceptual Captions 12M (CC12M) dataset.

Collection Method

Images were encoded into feature representations using a VQGAN f16 1024 model.

Freshness

Last updated on 2022-04-23.