kiara-01-lab/fast-manga-oss

FastManga OSS, also called FDS for FastLoop Documentary System, is a Python tool that reads activity logs from work tools (Slack, Notion, GitHub) and produces a four-panel manga strip telling a human story about what the team did. The pitch in the README is that small contributors who do real work but cannot self-promote are usually invisible, and this tool finds them in the logs and tells their story automatically. Three example stories ship with the repo: a Tokyo startup founder pitching with no time, a Taipei fashion brand pop-up, and an NPO running a zero-budget marketing push. The pipeline has five stages. The first stage scores activity logs to pick a protagonist, and is fully deterministic so the result is reproducible. The second stage calls a large language model to write a four-act story using the classical Japanese Ki-Sho-Ten-Ketsu structure (setup, escalation, turning point, resolution). The third stage compiles that story into a panel layout, again deterministic and unit-tested. The fourth stage generates the panel images, with a pluggable adapter that defaults to placeholders and can switch to DALL-E or Stable Diffusion. The fifth optional stage adds audio through MusicGen or Coqui TTS. The default run needs no API keys: it produces placeholder images and serves as a smoke test for the rest. To get real stories, the README documents four options: a local Ollama setup with Qwen3.5 72B, an Anthropic Claude cloud setup, DALL-E for real images, or a local Stable Diffusion server through Automatic1111. Users write a brief as a JSON file with fields for protagonist, problem, the role the product plays, the resolution, the emotion, the setting, and an optional activity log. Three deployment tiers let the user choose whether Chinese LLMs are included, excluded, or whether only local models are used, which the README frames as a compliance choice for regulated environments. The project also includes a prompt sanitizer that blocks named manga and anime IP before generation, replacing the term with a generic equivalent, and a provenance logger that records the original and sanitized prompt, the model, and the seed for each image. The code is MIT licensed.