""" Image decoding. In the new architecture, each detector does its own tensor preprocessing (CLIP needs CLIP's specific resize/normalise; frequency detectors don't resize at all). This module just turns raw bytes into a validated PIL Image that all detectors can consume. """ from __future__ import annotations from io import BytesIO from PIL import Image, UnidentifiedImageError # Pillow's default decompression-bomb guard is ~89 MP. Tighten it: a 25 MP # image is more than enough for any consumer photo, and rejecting larger # inputs caps memory usage on the inference server. Image.MAX_IMAGE_PIXELS = 25_000_000 class InvalidImageError(ValueError): """Raised when the bytes do not decode as a usable image.""" def decode(image_bytes: bytes) -> Image.Image: """Decode raw bytes into an RGB PIL image. Raises InvalidImageError on failure.""" try: image = Image.open(BytesIO(image_bytes)) image.load() # Force decode now so we catch errors here, not later. except (UnidentifiedImageError, OSError) as exc: raise InvalidImageError(f"Could not decode image: {exc}") from exc except Image.DecompressionBombError as exc: raise InvalidImageError(f"Image too large: {exc}") from exc if image.mode != "RGB": image = image.convert("RGB") return image