Skip to content

Cast Overview

In comic and manhwa recaps, voice narration and dialogue attribution depend entirely on knowing who is speaking and who is on screen. Because comic pages do not have built-in script labels, Cast transforms raw panel scans into a structured, named character roster linked directly to your story’s dialogue.

Cast workspace showing the chapter list on the left, character groups in the center, and unassigned clusters on the right
The Cast workspace: chapter selection, character clustering, and group management.

Manhwa artwork presents unique challenges for automated recap generation:

  • Visual transformations: Characters undergo art style evolutions, wardrobe changes, power-ups, injuries, or time skips between chapters.
  • Detached naming: A character’s name is often spoken chapters before or after their first physical appearance.
  • Ambiguous speech bubbles: Dialogue balloons point to off-screen characters, narration boxes float over crowds, and multi-speaker panels require visual spatial reasoning.

To keep results accurate and easy to edit, RecapSynth breaks character identity into five progressive steps:

Stage 1
Head Detection

Detect and crop character faces inside panel bounds.

Stage 2
Chapter Clustering

Group faces within each chapter so each character has a local cluster.

Stage 3
Character Grouping

Merge matching chapter clusters across the whole manhwa into Character Groups.

Stage 4
Name Assignment

Extract names from dialogue and match them visually to your groups.

Stage 5
Script Assignment

VLM attributes each dialogue balloon and writes panel action descriptions.


  1. 1

    Head Detection

    Open Guide

    Scans inside detected panel (art) bounding boxes to crop character faces and assign initial coordinates. Includes a manual canvas editor to draw missed faces, reposition boxes, or adjust crop boundaries.

  2. 2

    Chapter Clustering

    Open Guide

    Operates on the Chapters tab. An AI facial recognition model extracts feature embeddings and groups similar faces together within a single chapter. Unsure or one-off faces land in Unassigned for manual drag-and-drop sorting.

  3. 3

    Character Grouping

    Open Guide

    Operates on the Characters tab. Rolls chapter clusters up across the entire project. Matches a protagonist’s appearances from Chapter 1 through Chapter 50 into a single unified Character Group, regardless of visual changes.

  4. 4

    Name Assignment

    Open Guide

    Connects names to faces. An LLM scans transcribed dialogue lines to identify character names and aliases, while a VLM inspects surrounding panels to suggest matches for your Character Groups. Includes an Alias Evidence Visualizer to verify dialogue context before approving.

  5. 5

    Script Assignment

    Open Guide

    A Vision-Language Model processes full page images alongside the page’s character roster. It identifies which character spoke each dialogue line, routes system captions to the Narrator, and generates action descriptions for every panel.


Every RecapSynth project automatically seeds two non-deletable system entities to guarantee that scripts never fail during generation:

Narrator

Receives rectangular text boxes, system status screens, skill notifications, time cards, and expository voiceovers.

Unknown Character

Acts as a safe fallback for off-screen voices, crowded background shouts, or dialogue whose speaker cannot be determined visually.


To begin setting up your cast, navigate to Cast in the primary navigation rail and start with Head Detection.

Start with Head Detection

Detect character face bounding boxes across your chapters and inspect crop accuracy on the canvas.

Open Head Detection