from __future__ import annotations from typing import Any from PIL import Image, ImageChops COPY_INSTRUCTION = " Copy the complete printed value verbatim, including every digit, punctuation and suffix. Do not shorten, sanitize or canonicalize it." def preserve_scalar_fields( fields: dict[str, dict[str, Any]], ) -> dict[str, dict[str, Any]]: return { name: {**spec, "instructions": spec.get("instructions", "") + COPY_INSTRUCTION} if spec.get("type") == "text" and (not spec.get("knowledge")) and isinstance(spec.get("instructions", ""), str) else spec for (name, spec) in fields.items() } def focus_document( image: Image.Image, *, max_pixels: int = 1600000, max_scale: float = 2.0 ) -> Image.Image: if image.mode != "RGB": return image bounds = ImageChops.difference( image, Image.new("RGB", image.size, "white") ).getbbox() if bounds is None: return image (left, top, right, bottom) = bounds bounds = ( max(0, left - 24), max(0, top - 24), min(image.width, right + 24), min(image.height, bottom + 24), ) (width, height) = (bounds[2] - bounds[0], bounds[3] - bounds[1]) if width * height > image.width * image.height / 2: return image scale = min(max_scale, (max_pixels / (width * height)) ** 0.5) if scale <= 1: return image focused = image.crop(bounds) return focused.resize( (max(1, int(width * scale)), max(1, int(height * scale))), Image.Resampling.BICUBIC, )