Using Chrome AI to Summarize Comic Books (www.raymondcamden.com)

A developer extended a web-based comic reader to run on-device multimodal AI using Chrome’s LanguageModel (Prompt API) and a Summarizer API to automatically summarize comic books. The pipeline unpacks .cbr/.cbz archives (zip.js / Unarchiver), exposes a binreader for raw image blobs, and feeds each page image to a Prompt API session configured with an image input schema and a system prompt that asks for concise page-level paragraphs. Page summaries (up to 50 pages) are collected and then passed to the Summarizer.summarize call to produce a single book-level TL;DR. The implementation includes UI progress updates, handles initial model downloads, and monitors/quota-manages the prompt session by cloning it when inputUsage/inputQuota exceeds ~75%. This demonstrates practical, privacy-friendly on-device multimodal summarization without cloud uploads, but highlights limits: per-page latency (~2–3s), context-window restrictions requiring session cloning, imperfect filtering (ads/boilerplate bleed into summaries), and quality that trails a cloud Gemini run. Key snippets include sending images as {type:"image", value: await binreader(pages[i])}, using a paragraphSchema to constrain output, and calling Summarizer.create({format:'plain-text', length:'long', type:'tldr'}). It’s a useful proof-of-concept for offline, client-side comic analysis that needs prompt, schema, and UI tuning to reach production-quality results.