Pillow

The essential Python library for resizing, padding, and normalizing image data to meet strict VLM architectural constraints.

Pillow manages the critical transformation layer between raw visual files and model-ready tensors. It standardizes disparate inputs into the fixed resolutions (often 224x224 or 336x336) required by architectures like CLIP or LLaVA. By leveraging high-quality resampling filters (such as Lanczos) and precise canvas padding, the library prevents aspect ratio distortion that degrades zero-shot performance. This ensures every pixel aligns perfectly with the spatial expectations of the Vision Transformer (ViT) backbone.

https://python-pillow.org/