MangaCut

by slashneck · Creative

Turns manga and comic pages into animated, voiced stories: panels, animation, voices and sound, all on your own PC.

✓ VerifiedOpen source · GPL-3.0Unsigned↓ 578 downloads★ 0 on GitHubNo ratings yetWindows

About

MangaCut

Latest release Downloads Windows 10 and 11 GPL 3.0

Download for Windows   |   Install   |   Privacy   |   Build it yourself


Turn manga and comic pages into animated, voiced stories - on your own PC.

MangaCut cuts pages into panels, removes speech bubbles and effects with an AI fill, hands panels to WAN (ComfyUI) to animate them, puts the animations back on the page and plays it all as a story: panels revealed one after the other, voices for every bubble, sound effects and music on a timeline. Finished stories render to MP4 or GIF.

  • Everything stays on your PC. MangaCut sends nothing anywhere: no account, no tracking, no telemetry.
  • Small to install. AI models, voices and sound packs are only downloaded when you press their button in the app, from fixed addresses, each file checked against its SHA-256.
  • Windows 10 / 11, 64-bit. An NVIDIA graphics card is recommended for the voice engines and for animating.

Install

Download MangaCutSetup-x.y.z.exe from the latest release and run it. It installs for your Windows user only (no admin rights) into %LOCALAPPDATA%\Programs\MangaCut, adds a Start menu shortcut and shows up in Installed apps, where it can be uninstalled again. The portable MangaCut-x.y.z-win-x64.zip works too: unzip it and run MangaCut.exe.

Windows SmartScreen may warn the first time because the files are not code-signed: More info → Run anyway. Each release lists the SHA-256 of its files.

Updates: Settings → Check for updates asks GitHub whether a newer version is out - that is the only time MangaCut goes online by itself, and only when you ask (or switch on Check for updates when MangaCut starts). An installed copy then downloads the new setup, checks it and updates itself; your projects, settings and downloads stay.

Optional parts (each one a button in the app): the AI background fill (LaMa, 200 MB), the voice engines Kokoro, IndexTTS and Chatterbox, the voice library, the vocal sound library, and - for the ANIMATE tab - your own ComfyUI (Pinokio, portable or desktop) with WAN 2.x models. The ANIMATE tab has a setup guide.

Projects

  • Click the project's name (top bar) to rename it - its file and folder are renamed.

  • File → Add pages (select several) or drop images on the window. Pages show on the left; drag to reorder, PgUp / PgDn to flip.

  • Ctrl+S saves a .mangacut project file. Masks, exports and animations go in the <name>_files folder next to it. Double-click the project file to open it (the exe registers the file type for your Windows user on first start; if you move the exe, start it once again).

  • Pages edited with the first MangaCut version (<page>_mangacut folders) are picked up automatically when added.

Panels

  • Auto-detect finds panels along white or black gutters, including slanted ones and gutters a character overlaps. Panel outlines sit just inside the black border, so the border is never part of a panel.
  • View: Page · Cut (everything outside the panels blacked out) · Clean (panels as WAN gets them).
  • Edit (V): drag a side to move it (it snaps onto the nearest border line; hold Alt to skip, Ctrl to move freely), drag a corner, drag inside to move, double-click a side to add a corner, right-click a corner to remove it.
  • Point (A) or Manual: click on a side to add a point and drag it into place · right-click a point removes it.
  • Shape (selected panels): Add points puts a point in the middle of every side, Simplify removes unneeded ones, Round rounds the corners (non-destructive, 0 = sharp again).
  • Something sticking out into another panel: draw the panel's shape around it (Manual points), tick On top of other panels - that part animates with its own panel and covers the neighbour only there.
  • Dragged sides snap parallel onto the inner edge of a border line only when released within a few pixels of it (highlighted while dragging, Alt = no snap).
  • Split (S) draws a line through a merged panel · shift-click several + Merge · Rect (R) / Poly (P) · Del.
  • Panel numbers are their order (1…n). Drag panels in the list to reorder, Sort = reading order. Ctrl+Z undoes.
  • Navigation: scroll = smooth zoom at the cursor · drag empty space / middle mouse / Space = pan · F = fit.

Story (STORY tab)

Plays the project like a storybook: each page starts hidden, panels are revealed in their number order (reorder with drag / Earlier / Later), play their animation and freeze on the last frame, then the next one comes. A finished page is shown in full and fades to the next page. Space = play / pause, ← → = panel, PgUp / PgDn = page, F = fullscreen, drag the bar to jump. Every setting updates the preview right away; hover any option for an explanation. Opening the tab makes every page ready once, one after the other (the MangaCut logo shows how far it is - Watch now starts right away, the rest goes on in the background), and keeps them: switching pages, leaving the tab and coming back is instant. Only a page that changed meanwhile is made again. If the whole story would not fit in memory, the pages around the one shown are kept and the others are made when you get there. Removing bubbles with the AI fill is the slow part the first time (about 1-2 s per bubble / effect group); fills done in the Clean view or the WAN export are reused, and all fills are kept in the project folder, so the next time is instant.

  • This panel: seconds, speed, how often it plays, speech mode - for the panel in the preview.
  • Speech: Typewriter - each bubble appears when it is spoken and its original lettering types in, bubble after bubble. Static - bubbles are just there. Clean - no bubbles. Typing speed, pauses, reading time, text direction (lines / Japanese columns), bubble order. Longer speech (an animation shorter than its panel's speech): Loop, then finish the loop (keeps looping, then plays to the loop's end) · Loop, freeze when said (keeps looping and stops on the frame it is at when the speech ends) · Play its plays, then freeze (plays as often as set, then stays on its last frame).
  • Timing, Look (hidden panels black / dimmed / blurred; borders and frames outside the panels appear with the nearest panel, are visible from the start, or at the page end), Render (MP4 / GIF; frame rate Smooth = a whole multiple of the animations' own speed, e.g. 32 fps for 16 fps animations, so every frame is shown equally long).

Script (below the preview)

Every item of the page marked Keep on top, in playing order, one row each: # (panel·item - click the row to see it in the preview, paused just before it appears; double-click opens the line), Type (Bubble - appears complete, then the text types in · Text - letters appear with their outline · Sound FX - fades in · Always · Hidden), Voice, How (Out loud · Thought · Whisper - cloud-shaped bubbles start as thoughts), Feeling (one of the voice's feelings with IndexTTS; the intensity 0 - 4 with Chatterbox / Kokoro), Words (+ puts a sound into them), Status (✓ ready · ● to make · making… · ♪ own), and ▶ hear (made first if needed) · ⟳ new take · ✎ the whole line · ♪ own audio · ↻ read the words again · ↑ ↓ order.

Editing never rebuilds the list or moves the preview to another panel. Make voices / All pages make every missing or changed line in the background - keep working meanwhile (Stop cancels the rest). A line changed while it is being made is made again.

How a line is said:

  • Out loud / Thought / Whisper. A thought is the same voice, softer and closer, with a little room around it (how much: Thought room in the voice). A whisper is breath instead of voice (better: a whispering feeling). These are effects put on the made take afterwards: switching a line to a thought (or changing the room) never makes it again - the performance stays exactly the same.

  • Feeling (IndexTTS): one of the voice's feelings - a clip that sounds that way (bored, breathless, excited…) or emotion amounts - and its strength (✎). The voice stays the voice's own; only the way it speaks is copied.

  • Intensity (Chatterbox / Kokoro): 0 - 4, from calm to overwhelmed - the line is broken into pieces with breaths between them, stammers, a shaky voice (effects - they cannot add real emotion to these engines).

  • Sounds in the words: [sigh], [gasp], [laugh], [breath], [whimper], [pause] or the name of one of the voice's sounds - with how, words in any order: soft / strong / peak (or gentle, hard, heavy, intense…), low / high, short / long - e.g. "I- I can't… [soft sigh] stop… [strong low gasp]", "[heavy breath]", "[long sigh]". The voice's own sounds of that kind are used (the right strength, the lower / higher ones, a little pitch shift), else the sound library's; with IndexTTS breathing is acted in the voice itself. A [sound] the voice has nothing for is marked ⚠ in the script. Chatterbox Turbo also acts [sigh] [laugh] [gasp]… itself. "—" cuts a word off, "..." trails off. A line that is only a sound (Hm…, Nngh…, Ahh…) becomes that sound.

  • ✎ The line: words, way of speaking, feeling and strength (or intensity), Performed as - the exact pieces - and Write it into the words. Fine tuning: which sample speaks it, speed, pitch, variety. Takes: ◀ ▶ step through takes, ⟳ a new one.

  • The take, seen (✎ the line, and Try it in the voice editor): its loudness (grey), the melody of the voice (white line - flat or lively), the pieces it is made of (words · [sounds] in amber) and, in plain words: pace (slow / normal / fast), melody (semitones: flat = bored, calm · lively), breath (clear · breathy), end (clean, or something after the words → ⟳). Hover a value for what changes it; click the take to hear it from there. Zoom: mouse wheel where the mouse is (Shift + wheel moves along, double-click shows all; zoomed in far, the wave itself is drawn). Trim: drag over a part, then ✂ Keep this part or ✂ Cut it out (or Delete); ↺ Undo trims. Trims are done to the made take - nothing is made again.

  • Shape its sounds (✎ the line, when it has sounds): length, pitch, bend (falls / rises), start (swells in / sharp), end (cut short / fades out), breath, trembling and loudness of the line's sounds - the words stay as they are.

  • Made speech is kept (%LOCALAPPDATA%\MangaCut\cache\speech): a line made again with the same words, voice and feeling comes out exactly the same - only what changed is new.

  • Every made line is checked: the engine can only talk as long as the words need, noise after the last word is cut off (a feeling's breath at the end stays), and a take that came out far too long, too short or cut off is made again - the best of three is kept.

  • Words are read from the page (Read words automatically): offline text recognition (PaddleOCR models inside RapidOCR, on this PC). ALL-CAPS lettering becomes normal sentences. Check the result - lettering can be misread. Horizontal lettering only (English and other Latin-script text); vertical Japanese columns are skipped.

  • ♪ Own audio: any item can use one of the voice's sounds or your own audio file (wav, mp3, ogg, flac, m4a…) instead of a voice. Same loudness as the voices; the text types along with it and the panel waits for it. A sound effect's sound starts as it fades in. The file browser plays a file as soon as you click it.

Voices (sidebar)

Explanations ? shows or hides every explanation in the voice screens (remembered).

Engines - each downloaded once when you ask, then fully offline (while making voices they cannot reach the internet):

  • IndexTTS-2.5 (bilibili IndexTeam; model free for non-commercial use) - acts feelings: the voice is copied from one sample, the way it speaks from another clip (or from amounts of eight emotions). Install IndexTTS… (≈ 6 GB engine: Python and PyTorch exactly as IndexTTS is tested, from GitHub / pytorch.org / PyPI, every package checked against its lock file; ≈ 7 GB of models from Hugging Face at fixed versions, every file checked). Best with an NVIDIA graphics card. English, Japanese, Chinese, Spanish, Arabic. In %LOCALAPPDATA%\MangaCut\engines\indextts. Speed: about as long as the line itself on an RTX 2070 (the first line after starting ≈ 30 s more). It needs about 6 GB of graphics memory while speaking - with WAN or a game holding the card it gets many times slower (shown while making voices). It lets go of the card after 10 minutes without speaking.
  • Other engines: Chatterbox (Resemble AI, MIT - natural, [sigh] / [laugh] tags; it puts an inaudible AI watermark into its audio) and Kokoro (small, any PC, flat).

No trained voice files are needed: a voice is copied from one clean clip of someone talking - from the character's clips, or from the voice library.

Voice library (Voice library… or New voice → From the voice library): hundreds of ready voices - click one to hear it, Listen through plays a list one after another. One woman, many feelings has the same woman in 23 moods (neutral, serene, cute, pain, distress, sad…) - with IndexTTS these make good feelings for any voice. Also acted styles (whisper, sad, happy, laughing…), studio voices, English accents and donated voices. From Kyutai's voice collection on Hugging Face (made for voice cloning), at a fixed version, every file checked; only anonymous voices. A voice is fetched when you first click it (≈ 0.5 - 3 MB) and kept in %LOCALAPPDATA%\MangaCut\voices\library. Licenses are shown (NC = non-commercial: fine for private projects).

The voice editor (every setting shows its number; ↺ goes back to the usual value; the window keeps its size):

  1. Voice - samples: From clips… (the voice maker), Voice library…, or More ▾ (voices made with Kokoro). With Chatterbox each sample is used for something: the normal voice, from an intensity on, thoughts, whispers.
  2. Feelings (IndexTTS) - New feeling…: from a clip (a file - click to hear it -, one of the voice's own sounds, or the voice library; drag over the waveform to use 3 - 15 s of it) or from emotions (eight amounts, with presets: calm, bored / resigned, weary, sad, afraid, shy, surprised, angry, happy…), a strength, and Try it. Usually: the feeling lines get unless they pick another.
  3. Sounds - sighs, gasps, whimpers, breathing… New sound in this voice… (IndexTTS): write the sound the way a bubble would - Hah! (short, sharp), Haah… (long, fading), Hnn… (held back), Hah… hah… (out of breath) - the palette puts in the usual letters; pick how it sounds (like the sound library's recordings of that kind and strength, the voice sample's own way, or one of the voice's feelings) and how much; Make 4 takes - IndexTTS acts it in this voice; tick the takes to keep, trim each, and shape them: length, pitch, bend (falls ↔ rises), start (swells in ↔ sharp), end (cut short ↔ fades out), breath, trembling, loudness (shaping never makes it again). Its kind and strength decide which [sigh], [strong gasp]… uses it. From files…: your own recordings. Sound library… (whose recordings this voice uses - automatically the woman nearest in height) and Use the sound library; Made - Acted in this voice (usual: IndexTTS acts the sound in this voice, copying how the recording sounds), Recorded (the recordings as they are - another person's voice) or Turned into this voice (experimental); Try sounds plays a soft and a strong sigh, a gasp, a whimper and breathing. A bubble that is only a sound (Hah!, Haah…, Hnn…) is acted from its own letters in this voice - like the voice's own sound of that kind (and shaped like it), else like the library's; '!' gives it a sharp start and a short end, '…' / '~' a fading one, '—' cuts it off.
  4. How it speaks - way of speaking, thought room, variety (steady ↔ lively: low for a flat, bored voice), speed, pitch, language (IndexTTS) / intensity and engine settings (Chatterbox / Kokoro). Try it: a sentence with any way of speaking, feeling or intensity, and sample.

Sound library (Sound library… in the sidebar or the voice editor) - real recorded vocal sounds: sighs, gasps, whimpers, cries, cheers and growls, each soft, medium, strong and peak, by eleven adult women: the VIVAE corpus (Holz, Larrouy-Maestri & Poeppel, Max Planck Institute for Empirical Aesthetics; CC BY-NC 4.0 - fine for private projects), recorded for emotion research. One download of 94 MB from Zenodo (CERN's research archive) when you ask, checked against its fixed SHA-256, unpacked to %LOCALAPPDATA%\MangaCut\sounds\vivae; Remove the recordings deletes them. Click a sound to hear it; Listen through.

Voice maker - drop clips of the character talking into it. Short clips are fine: add several with the same energy and they are joined into one sample (breath and noise around the talking are cut off, so the voice does not learn to breathe out after every line); if it is still short it is repeated up to what the engine needs. Only that voice - no music, effects or other voices. Untick a clip to leave it out; click a clip and drag over its waveform to use a part. MangaCut only reads audio and Kokoro voice numbers - it never runs anything from a downloaded file.

Kokoro file… imports community Kokoro voices (.pt / .npy / .npz) safely; Export voice clips (MP3) saves every line.

Batch (all panels at once)

  1. Export all (or Export page) writes every cleaned panel into one folder, named mc001-01_…png (page-panel).
  2. Point ComfyUI's batch image loader at that folder and run WAN.
  3. Import animations… and pick WAN's output folder. Files are matched by name if they still contain the mc001-01 tag, otherwise by order. A review list shows each panel's input next to its GIF before importing. If WAN saved a GIF and an MP4 with the same name, the GIF is used.

Overlapping panels (a leg reaching into the next panel)

Draw the panel's shape loosely around the part that sticks out (Manual / Add points) - no need to trace it exactly, and no need to cut it out of the other panel. MangaCut finds the part itself inside your outline (its lines and everything they enclose - not the other panel's background you caught with it) and shows it hatched in the PAGE tab; a cut-out in the other panel shows dotted. The sidebar says which panel it reaches into.

  • The other panel's WAN input gets its own background where the part lies (filled in) - no white shape for WAN. Everything of the other panel inside your outline stays that panel's: it appears only when that panel does.
  • This panel goes to WAN with room around the part, so it can move.
  • Afterwards the part is drawn on top of the other panel. Move with the animation (default): frame by frame only what the animation draws there is kept - the part moves freely and is not cut off by its old outline. Stay still: it keeps its place. Off: treated like the rest of the shape. Room to move sets how far it may go. Its shape: Exactly my outline (usual - everything inside the shape you drew is this panel's part) or Found inside a rough outline (for a loosely drawn shape - background caught in it stays with the other panel). Export again after changing it. A moving part keeps everything of it that the animation still shows as drawn - white inside a shoe or a sock is the part, not the white around it - even where it touches a panel's edge.
  • Which part of a shape sticks out is worked out from the panel's frame lines - the long straight edges of the shape, however big the leg and shoe sticking out are; a panel cut around another panel's part keeps its body (the cut is a cut-out for the part, not a slanted side).
  • Where the part was, the page shows what is behind it (filled in, panel borders drawn on straight), so a part that moves leaves no copy behind - in the page preview, renders and the story (there it appears together with its own panel).
  • Panel borders running across the part (a leg drawn over the divider) are taken off it before it goes to WAN; its own lines that cross them carry on. A border the part only rests on (a sole standing on the panel's bottom edge) stays - it is the part's outline there. Its whole outline goes with it - also where your outline cuts through it.
  • The moving part is followed from the still picture into every frame of the animation (its shape as you drew it, moved the way the animation moves it); only where it changes too much to follow is it cut out by its outline. Its soft edge blends into the panel behind it (no white rim).
  • What lies inside a panel's mask is that panel's: another panel's animation never paints there (its WAN picture still gets its own background there), and while that other panel is hidden its hidden look does not reach into the part or across it.

Panel editor (double-click a panel)

Two kinds of marks:

Mode Use for In WAN input In final GIF
Keep on top (K), solid speech bubbles, text filled with background original put back on top, still
Remove for good (X), striped motion lines, effects filled with background stays removed, the animation shows

Tools: Wand (W) click inside a bubble · Brush (B), [ ] size · Lasso (L) · Switch (S) click a mark to change your mind: back on top ⇄ gone for good · right-click removes · Ctrl+Z undo · Auto bubbles · Convert all marks switches every mark in the panel to one mode.

Background fill guesses what is behind the marks. Clean view (2) shows exactly what WAN gets.

  • AI (LaMa): best. One-time ~200 MB download from the panel sidebar.
  • Texture: rebuilds screentone / patterns (default without the AI model).
  • Smooth / White: simple fills.

Then Export image (or Copy to clipboard → Ctrl+V on ComfyUI's Load Image node) at the WAN size (top right, default 384 × 576), generate, and Import GIF / MP4 (or drop it on the panel).

Animation options: Put bubbles back on top · Match tone to page (fixes WAN brightness drift) · Grayscale · Still area around bubbles (soft still border around kept bubbles, hides wobble) · Keep panel border.

Render the panel or the page (Play page previews) as GIF or MP4.

Refine (F) fixes the cleaned picture, spot by spot (right-click takes one back, Undo refine / Clear):

  • Remove: the painted spot is filled again (AI again, paper / flat tone, texture from around, smooth).
  • Improve: keeps what is there and makes it better (clean up, sharpen lines, smooth out noise) - with a strength.
  • Redraw: a picture model in your ComfyUI draws the spot new, fitting what is around it - the top of a head a bubble covered, a hand, hair. Say what belongs there, paint over the spot (a little beyond it). Again tries once more. Uses an SDXL / SD 1.5 checkpoint from ComfyUI (an anime one works best); the result is kept with the page.
  • Draw: by hand - Pencil (hard line, ink 0 = black … 255 = white), Brush (soft), White (paint paper), Clone (Alt+click where to copy from, then paint where it goes; a cross shows what is copied). Alt+click with Pencil / Brush picks the ink from the picture. Opacity and softness; the size number shows the brush in pixels.

The part of an animation used (Animation section, also in STORY → This panel): Start at, Length (All = to its end) and Loop: Plain · Smooth (its end blended into its start - no jump) · Back and forth. A 9 s loop of which only 2 s are needed: Length 2.0. Used everywhere (page, story, renders).

ANIMATE tab - WAN inside MangaCut

MangaCut runs WAN through ComfyUI on this PC with its own workflow (ComfyUI's built-in nodes only - nothing to wire). Only 127.0.0.1 is ever talked to.

Setup (the Setup guide button checks each step):

  1. ComfyUI - easiest via Pinokio (Discover → ComfyUI, one click). MangaCut finds Pinokio's ComfyUI by itself (or My ComfyUI folder…), and starts it when you press Generate (Settings → Animate).
  2. WAN models in ComfyUI's models folder (official, free: Hugging Face Comfy-Org/Wan_2.2_ComfyUI_Repackaged): diffusion_models: a WAN 2.2 I2V high noise + low noise pair (or one WAN 2.1 I2V model) · text_encoders: umt5_xxl_fp8_e4m3fn_scaled · vae: wan_2.1_vae · optional frame_interpolation: rife_v4.26 for Smoothing. Merged "Lightning / Lightspeed" models: 4 steps, guidance 1 (the presets). Plain WAN models: Fine tuning → 20 steps (10 motion), guidance 3.5 - or add a speed LoRA (lightx2v).
  3. 8 GB cards: Fast / Balanced, 2 - 3 s. MangaCut frees the voices' graphics memory first and keeps some graphics memory free when it starts ComfyUI (Settings) - a full card makes Windows spill into RAM and gets very slow.

Use: PANEL tab → Animate here ▸ (or PAGE tab → Animate all panels here) sends the cleaned, fitted panel; Add pictures… / dropping pictures animates any picture. Per picture: the motion prompt, length, ending (Loop - ends where it starts · start → end picture · free), variations (several seeds at once), seed. Shared: Quality (Draft · Fast · Balanced · Quality · Best - size and steps), Smoothing (RIFE ×2 … ×4), Models, LoRAs (pairs: high file on Motion, low file on Detail; ↑ ↓ change their order; Edit keeps trigger words and notes per LoRA file for every project - press a trigger word to put it into the prompt, again to take it out), Fine tuning (steps, motion steps, guidance, shift, sampler, scheduler, size, the words to avoid), When it is done. A live preview and progress show while it works; results play in the middle - Use on its panel puts one on the panel it came from (the first one is put on by itself), Save as… MP4 / GIF. Input folder / Output folder (below the list) open this project's pictures and animations - their arrow opens ComfyUI's own input / output folder.

Sound & music (STORY tab → SOUND)

Under the story preview, switch SCRIPT | SOUND. SOUND is a timeline of the whole story: the pages and panels on top, where voices speak, and tracks for sound effects and music.

  • + Sound… / + Music… (or double-click an empty spot of a track) open the sound library; Add puts the sound on a track at the playhead. Music loops until the story ends - drag its end to stop it where you want. Audio files can also be dropped straight onto a track.
  • Drag a sound to move it (also to another track) · drag its left / right edge to cut it shorter · drag the little handles on its top corners for fades · ✂ Cut (S) splits it at the playhead · Ctrl+D duplicates · Del removes · right-click for more. Edges snap to the playhead, pages, panels and other sounds (Alt = freely).
  • Click the ruler or an empty spot to put the playhead there (the preview jumps along); Space plays. Wheel: up / down through the tracks, Shift+wheel: left / right, Ctrl+wheel: zoom.
  • Several at once: Ctrl+click / Shift+click, or drag a box on an empty spot - selected sounds move together. Del removes · Ctrl+C / Ctrl+X / Ctrl+V copy, cut out, paste at the playhead · Ctrl+D duplicates · Ctrl+A · Esc · Ctrl+Z / Ctrl+Y undo / redo · S cut · M mute · L loop · ← → nudge (Shift: 1 s) · ↑ ↓ other track · Home / End · F the whole story. Keys in the bar lists them all.
  • On the right, the selected sound: sticks to the story's time, its page or its panel (then it moves along when that page / panel gets longer through speech), start, length, loop, loudness, fades, pitch, speed, left / right, quieter under voices, repeat N times every X s (each a little different), mute. Loudness of everything: music, effects, voices - sliders follow how loud it sounds (50 % = about half as loud).
  • Music plays on across page changes - only where its clip ends (with its fade) does it stop. Renders mix everything in stereo, 48 kHz.

The sound library (+ Sound…): Make - an AI sound model in your ComfyUI makes it from a few words: Sound effect (Stable Audio 3 small SFX, 3.5 GB) or Music (ACE-Step 1.5 turbo, 10 GB, any length, tempo and key) - each a one-time download (button in the tab) into ComfyUI's model folders, from Hugging Face, checked. My sounds - your own files (WAV, MP3, OGG, FLAC, M4A…).

Voices - more control

In the Line window: right-click a part of the take (a word piece, a [gasp]) → Make … again re-makes only that part. Make takes to choose from: several takes to hear side by side - ✓ uses one. IndexTTS: Steadiness (cleaner, steadier takes), Variety, and Its own feeling - a mix of eight emotions for this line only. Shape its sounds → Follow the voice: its sounds (gasps, sighs…) are brought to the voice's own pitch (a little higher the stronger they are). Thought voices: a clean room is added after the line is made (the voice itself stays clean) - Voice → Room size, Room tone, Echo, Close voice.

Settings (⚙, top right)

Where downloads are kept - voice engines, picture AI, voices, sounds, cache - Change folder… picks another folder / drive for it (also before anything is downloaded: new downloads then go there); what is there is moved (the same drive: just renamed, instant; another drive: copied many files at once, checked - same number of files, same size - and only then the old copy deleted; MangaCut restarts). About & updates: the version, Check for updates, and whether to check at each start (off by default). The ComfyUI link (folder, address, start it by itself, faster math, live previews, graphics memory kept free). Scrolling over a setting never changes it unless you clicked it first; numbers can be dragged left / right (Shift finer, Ctrl coarser) or clicked to type.

Packing a finished project

Pack (top bar, next to Save) or File → Pack project… writes one ordinary .zip with everything the project uses (page pictures, masks, refinements, the animations on its panels, the voice lines it speaks, voice samples, placed sounds / music, used ANIMATE results) and nothing else (old takes, unused variations, caches). It is read back to check it. File → Unpack a packed project… (or drop the zip on MangaCut) opens it again anywhere - or unzip it yourself and open the .mangacut file inside.

Privacy

  • No account, no analytics, no crash reports, no update checks unless you ask for them.
  • Online only when you press a download button (or Check for updates): Hugging Face, GitHub, pytorch.org, PyPI (the voice engines' Python libraries) and Zenodo - fixed versions, every model file checked against its SHA-256.
  • ComfyUI and the voice engines are only talked to on this PC (127.0.0.1). The voice engines run with Hugging Face's telemetry switched off and offline mode on.
  • Your pages, projects and recordings never leave your PC.

Building from source

Needs Python 3.12 and, for the installer, the .NET 8 SDK.

.\build.ps1            # dist\MangaCut.exe (sets up .venv from requirements.txt the first time)
.\build-release.ps1    # release\: setup, portable zip, checksums and source zip for a GitHub release

Or run it from source: .venv\Scripts\python.exe main.py. The version is __version__ in mangacut\__init__.py; publish a release under the tag v<version> with the files from release\ - the updater looks for MangaCutSetup-<version>.exe and its .sha256.

License

MangaCut is free software under the GNU General Public License version 3 (see LICENSE). It includes and builds on other projects - see THIRD-PARTY-NOTICES.md. The models, voices and sounds it can download have their own licenses (several are non-commercial only) - listed in THIRD-PARTY-NOTICES.md.

Patch notes

MangaCut 1.0.0

16 Nov 2026

The first public release of MangaCut: manga and comic pages turned into animated, voiced stories, all on your own PC.

Download: MangaCutSetup-1.0.0.exe installs for your Windows user (no admin rights) and can be uninstalled from Installed apps. Prefer no install? Unzip MangaCut-1.0.0-win-x64.zip and run MangaCut.exe.

  • Cut pages into panels (auto-detect, slanted gutters, parts that reach into the next panel)
  • Remove bubbles and effects with an AI fill, export panels for WAN, put the animations back
  • ANIMATE tab: WAN through your own ComfyUI, with a setup guide
  • STORY tab: panels revealed one by one, voices for every bubble (Kokoro, IndexTTS, Chatterbox), a sound and music timeline, MP4 / GIF render
  • Nothing is sent anywhere. AI models, voices and sounds are downloaded only when you press their button, each file checked against its SHA-256
  • Settings → Check for updates updates an installed copy from here, only when you ask

Windows SmartScreen may warn because the files are not code-signed: More info → Run anyway.

SHA-256

4a6ef56314f20d51006e033a2ce83308afaea39b617336a49f56d14eb7ab2d2a  MangaCutSetup-1.0.0.exe
523170d949f397b4ae4d3ad0123c0ba84c17aef70de1f88b4804629be4443ac3  MangaCut-1.0.0-win-x64.zip
All releases on GitHub ↗

Ratings & comments

No ratings yet. Be the first to rate it.

Sign in to rate:

No comments yet. Be the first to say how it ran.