One project, six stages
Source, Art, Story, Video, Sound, Deliver. Every stage reads what the one before it made, so nothing is re-described and nothing drifts.
Everything a page needs to become a film.
Shotink is a production studio, not a prompt box. Nineteen steps across six stages take a comic or manga page from ink to a finished, voiced, scored anime scene — and every one of them can be run on its own, fixed by hand, and run again.
Plan first, then generate
Each stage opens with a planning pass that reads your pages and writes its notes down. The generation steps read those notes instead of guessing.
Fix by hand, on the canvas
Paint a region and describe the repair, crop a panel, clone a background, re-say a line, re-run a shot. The result replaces the file in place, undoable.
A GPU of your own
Heavy steps rent a dedicated GPU for your job, with your model cache on a private volume. You see the estimate before you run and pay for the minutes used.
The six stages
Every step, what it needs, and what it hands on.
Steps are listed in the order you would normally run them. GPU steps rent a card for your job; AI steps run on the server against a language model and cost a few credits; free steps cost nothing.
STAGE 01
Source
Bring the comic in and break each page down into single panels, with the speech bubbles removed.
Import comic pages free
Drop in pages as images, a folder or a zip. Pages are kept in reading order and nothing is altered.
Understand Pages AI
Reads each whole page: transcribes every bubble, caption and sound effect, describes the scene, and reads the gutters — the action the reader infers between panels. Afterwards the assistant can discuss your comic from these notes instead of re-reading the images.
pages→one analysis per page
Split into Panels GPU
A trained ensemble of panel detectors crops every panel from the page in reading order, combining several models so a border one misses another catches.
pages→one image per panel
SelectPaintCrop
Find Speech Bubbles GPU
A second ensemble finds every balloon and knocks it out to transparency, leaving a see-through hole where the lettering was. Paint over any it missed.
panels→panels with bubbles knocked out
SelectPaintCrop
STAGE 02
Art
Repair the panels, repaint them in an anime style, and shape them ready for video.
Plan Art AI
Looks at each panel as a picture: who is in it, how it is framed, its own light and colour, and what has to be repaired before it can be a frame of film. It also decides whether a panel is a frame at all — a sliver of border or a sound effect over blank ground is marked not animatable, and every later step skips it. One click overrules it.
panels→a per-panel analysis and a clean-art prompt
Clean Up Panels GPU
Flux.1 Fill paints artwork back into the holes the bubbles left — one hole at a time, each from a crop around it and its own sentence from the panel notes, so what belongs behind one hole cannot end up in another.
panels with holes→clean panels at their original size
SelectGen FillClone
Apply Anime Style GPU
Restyles every panel into one of 60+ looks — studio anime, watercolour, painted, ink, stop-motion, classic manga masters — on Flux.2 Klein. The style is how the picture is drawn; each panel's own light and mood ride separately, so a cheerful page stays cheerful in a dark-action style. It is also the last step that can remove a border or a drawn-in sound effect.
clean panels→panels in the chosen style
SelectTry a Style
Fit to Frame GPU
Extends each plate outward until it is exactly the canvas the video model generates at, so every shot starts from the same aspect ratio. It only ever adds a margin: your plate is composited back over the result, so the picture itself cannot change.
styled panels→plates at video size
SelectGen FillClone
STAGE 03
Story
Read the pages as a story, then write the one script the film is made from — a narration or a dialogue, both cued.
Plan Story AI
Reads your imported pages — the originals, where the reading order and the whole moment are — and writes one story analysis each: what the page does, and every line said on it with a description of who says it. Descriptions survive restyling where a name the comic never printed would not.
pages→one story analysis per page, with beats
Write Script AI
Arranges what Plan Story read into one .srt you can open and edit. Narration is one voice; dialogue has the characters speaking, each cue naming who. Switch a finished script between the two without touching its timings. With nothing analysed it writes from your prompt; with nothing at all it hands you a timed outline to write over.
story analysis or a prompt→the film's script
Script editor
STAGE 04
Video
Cut a film out of the finished plates: plan the shots against the story, then generate them.
Plan Video AI
The only step that reads every plate at once. It decides which panels are worth a shot and drops the rest with a reason, groups consecutive panels that are one continuous moment into a single shot, fixes the film's lighting, palette and cast once so a jacket cannot change colour between cuts, and — where you wrote a script — sizes each shot to the line spoken over it. Everything lands on a storyboard you can edit.
plates + script→the storyboard
Generate Video GPU
MiniMax H3 renders each shot as one continuous take, with your plate pinned as the first frame so the shot starts exactly on your art. Every plate in a shot goes to the model as reference art; the storyboard's direction does the rest, including the gutter — the action the comic never drew. Re-run one shot with a better description and it replaces the clip in place.
storyboard→one clip per shot, 1344×768
Fix this shotTrim
STAGE 05
Sound
Plan the sound, design the voices, then voice it, score it and clean it — in that order, so the last step makes it sound designed rather than assembled.
Plan Sound free
Takes the clips and the script and casts a voice for every speaking character — the same voice all film, and never two characters sharing one in a scene. Fits each line to the clip it is spoken over and flags the ones that do not fit, which is a shot to lengthen rather than a line to rush. Cues a mood per scene for the effects pass.
clips + script→a cast, a reading speed per line, a mood per scene
Design Voices GPU
Every speaker gets a one-line voice description from the plan — rewrite it in your own words — and this renders each one into a reference clip: an audition take in that voice. Or upload a sample of a voice you have the right to use. From then on the reading is cloned from the clip, so a character sounds the same across the film and across re-takes.
voice descriptions→one reference clip per speaker
Generate Voice-over GPU
Qwen3-TTS reads the script aloud, one file per cue: a preset voice per speaker, or a clone of the speaker's reference clip, at the rate the plan fitted to each shot. A line is a file, so a bad reading is fixed by saying that one line again.
script + cast→one voice file per cue
Re-say a line
Generate SFX GPU
MMAudio watches each clip and lays scene-synced ambience and impact sound beneath it, prompted by the mood Plan Sound cued. A film with dialogue and no room tone sounds unfinished.
clips + mood→one effects track per clip
Clean Sound GPU
A mastering pass over every audio clip of the film: hiss and artefacts out, levels matched, so voices and effects sit together cleanly.
audio→mastered audio
STAGE 06
Deliver
Combine everything into the one file you take away. Nothing is baked in until you ask for it, so a clip or a track changed upstream is picked up by the next render.
Cut free
Lay the generated clips out in order with their transitions. The head and tail of every clip are trimmed automatically, and the cut is laid fresh each time from whatever the clips are now.
Mixdown free
Every voice-over at its cue and every effect under the shot it was made for, balanced and loudness-normalised to a broadcast target. Play it against the picture before you render.
voice + effects→one mixed soundtrack
Render & Export free upscale optional
Joins picture and soundtrack into one finished video, burns in the film's own script as subtitles if you ask, and optionally upscales the finished film with ESRGAN — once, on the final cut, rather than on twenty separate clips. Export the film alone, or the whole project as a zip with a manifest.
cut + mix + script→the finished anime
On the canvas
Six tools, each on the step where it is needed.
A tool appears on the rail of the step whose mistakes it fixes, and nowhere else — so the rail is short enough to be read.
Select
The cursor back to normal, and the way to pick a region for the other tools.
Paint
One brush whose meaning is decided by where you use it: on a panel it knocks out a balloon the detector missed; on a plate it marks the region to redraw.
Crop
Cut a border or a neighbour out of a panel the splitter took too much of.
Generative Fill
Paint over what came out wrong and say what should be there in plain words. Two models to choose from — an instruction editor that can hold “give the hand five fingers”, and an inpainter for pure repairs.
Clone
Copy a region of the image over another. When the right answer is “more of the wall six inches to the left”, copying beats imagining. No GPU, no queue.
Try a Style
One panel, one look, in place — before committing a folder of forty-nine to it. It shares the picker with Apply Anime Style.
The editors
Every file the studio makes can be opened and fixed.
Image editor
Pages, panels and plates with the tool rail for that step, a per-file edit history you can walk back, and the file's own notes beside it.
Script editor
The film's .srt as cues you can rewrite, retime and reorder. Switch narration to dialogue or back; mark the script the film is planned against.
Clip editor
Watch a shot, describe what is wrong, and render it again. Trim the head and tail. The new take replaces the old in place, undoable.
Voice editor
One file per line, so one bad reading is one line to say again — in the same voice, at the same cue.
Help that knows your project
An assistant that has read your pages, and a director that can run a stage.
The project assistant
Chat about the comic from the notes the planning passes wrote — who is in a panel, what a page does, why a shot was dropped. It can read any file's insights and propose the next step; it never spends a GPU without you.
The director
Point it at a stage, or a run of stages, and it works through the steps end to end: inspect, choose the next step, run it, judge it, repeat. Sample three panels before committing two hundred, park at each stage boundary, stop at any time. Every guard is enforced in code, not in a prompt.
Under the hood
Best-in-class open models, each doing the one job it is good at.
Nothing here is a black box: every step names its model and its settings, and you can change either.
Panel & bubble detectorsTrained YOLO ensembles, combined across models
Flux.1 FillCleanup, Fit to Frame and inpainting repairs
Flux.2 Klein 9BAnime restyling and instruction-based Generative Fill
MiniMax H3Image-to-video, one continuous take per shot
Qwen3-TTSVoice-over: nine presets, cloning from a clip, voices designed from a description
MMAudioScene-synced sound effects and ambience
ESRGANFrame-by-frame upscale of the finished film
Vision-language planningPage, panel, story, shot and sound planning
See it on your own pages.
Create an account, import a chapter, and run the planning passes for a few credits before you spend a minute of GPU.