Image Tiler

概要

Many LLM vision systems automatically downscale large input images, leading to significant loss of detail and rendering small text unreadable. This tool functions as an MCP server that intelligently tiles expansive images and full-page screenshots into smaller, model-optimized segments. It calculates an optimal grid based on the target LLM's vision configuration, extracts each tile, and provides comprehensive metadata including token estimates, ensuring that AI vision models can process every critical detail at full resolution without downscaling.

主な機能

• 1 GitHub stars
• Offers customizable tiling parameters, including target model, tile size, output format (WebP/PNG), and image pre-downscaling.
• Captures full-page screenshots from URLs, including scroll-stitching for pages over 16,384px.
• Prevents LLM vision downscaling by intelligently tiling large images and screenshots.
• Provides paginated retrieval of image tiles, complete with content hints and brightness statistics for each tile.
• Compares and estimates tile counts and vision tokens for various LLM models (Claude, OpenAI, Gemini).

ユースケース

• Ensuring LLMs can accurately read and interpret small text within large images or extensive web page captures.
• Analyzing highly detailed UI designs, complex diagrams, or dense documents with LLM vision models.
• Processing comprehensive web page content for AI analysis, overcoming typical screenshot resolution limitations.