Manga pages in. Voiced video out.

From manga page to anime episode

manga anime is the last step. Each panel is repainted as a clean key image, animated with its sound effects for as long as its shot lasts, and cut full-frame with the voices and subtitles. Page four of this demo, from the same analysis and voices as the video above. Turn the sound on:

# what would this cost? no model is called
manga anime bj --backend openrouter \
  --image-model black-forest-labs/flux-3-image \
  --video-model alibaba/wan-3.0 --dry-run

# check the images before paying for clips
manga anime bj … --keyframes-only

# the episode, refused above 14 s of video
manga anime bj … --key-shots-only --max-seconds 18

AI-colorized and animated adaptation; the colors and sound effects are invented by the models. Redrawing five images during tuning brought the real spend to $2.97. Mouths follow timing cues in the prompt, which is not exact lip sync, and there is no music yet.

Sound effects, one look, the same faces

What makes it feel like one episode instead of a series of clips.

# the sounds printed on the page
manga animate bj p0004_k0 \
  --sound "knuckles knocking on a door" \
  --characters Saito

# the end of a run
1 shots without sound effects:
  p0004_k1: still image, no clip. To fix:
  manga animate bj p0004_k1 --sound "..."

# the next episode, same look and faces
manga anime bj2 … --style-from bj

Seven commands from archive to video

Every step is cached and can be resumed. The analysis and the edit are plain JSON, so you can correct a line or retime a shot and render again.

# CBZ, EPUB, PDF, CBR or a folder of images
manga import enbj001.pdf --title "Black Jack"
manga project new bj --title "Black Jack" \
  --from ch1:p11 --to ch1:p14 --lang en
manga analyze bj   # panels, bubbles, OCR, speakers
manga read bj      # web reader to check and fix it
manga voice bj     # one voice per character
manga timeline bj  # the edit, as timeline.json
manga render bj    # out/bj.mp4

{
  "id": "0002_t4",
  "bbox": [320, 1020, 530, 1240],
  "panel": 2,
  "text": "Graduated from Eiroku University's medical department.",
  "speaker": "Director",
  "essential": true
}

Real pages need a second read

On page two of this demo, the local OCR saw two neighbouring bubbles as one. A review splits them, restores the words and names who is talking. Choose how that review happens below.

# Apple Vision, one box
"GRADUATED FROM EIROKU UNIVERSITY'S MEDICAL THIS WILL
DEPARTBE YOUR MENT. FIRST TIME ON DUTY... WELL..."

# after review, two bubbles
Director: "Graduated from Eiroku University's medical department."
Director: "This will be your first time on duty... well..."

Pick any model with your own keys

Search current models and prices without leaving the terminal, then pass the id to any step. Keys come from environment variables and are never written to disk.

$ manga models search image flux
bfl/flux-3-image-klein   gateway   $0.008-0.036/image
bfl/flux-kontext-pro     gateway   $0.04/image
bfl/flux-pro-1.1         gateway   $0.04/image
bfl/flux-3-image         gateway   $0.041-0.607/image

This demo ran on OpenRouter and Anthropic. The other providers are covered by automated tests only, through llm-models and llm-connect.

Everything manga-anim can do

New in this version:

• manga anime: a full-frame episode from an analyzed and voiced project, in color or black and white.
• Sound effects generated with the picture, and a report of the shots that have none.
• A style bible, a reference image and remembered faces, across runs and episodes.
• Vertical output with a maximum length and numbered parts, for TikTok, Shorts and Reels.
• A vision model can write each shot's motion prompt, with --direct-provider and --direct-model.
• --dry-run, --keyframes-only and --max-seconds keep the spending under control.
• The Claude Code skill installs with npx skills.
• Clips are no longer billed one second too long.
• Mouths move on talking shots, and the model no longer voices the line under the dub.

The full list:

Local by default, OpenRouter when you want it

The tool needs macOS on Apple Silicon, with ffmpeg, uv and Rust to build. Each generation step runs on your Mac or through OpenRouter with --backend openrouter, which is billed and sends your text and panels to third parties.