How AOT Beyond Is Made
What the process is—and is not
AOT Beyond is not traditionally hand-drawn manga, and I do not present it as such.
The underlying artwork is generated with AI. At the same time, the finished project is not the direct output of a single prompt or automated pipeline.
It is built through story development, art direction, visual experimentation, page planning, repeated generation, rejection, selection, continuity management, assembly, lettering, manual corrections, cleanup, and ongoing refinement.
The process is still evolving because the project itself is teaching me what works, what does not, and what needs to be controlled more carefully.
The goal is not to pretend the production method is something it is not. The goal is to use the tools available to turn a specific alternate ending into a coherent manga experience, while being honest about how that result was produced.
AOT Beyond uses an AI-assisted production process, but the actual workflow is considerably more involved than generating a few images and placing them on a page.
The story, overall direction, major twists, ending, page structure, dialogue, pacing, and emotional intent are developed by me. AI is used throughout parts of the process as a tool for brainstorming, challenging ideas, identifying weak points, and helping me iterate more quickly, but the project is built around a direction I already had and decisions I continue to make myself.
The artwork is AI-generated, although defining how the project should look has been an ongoing part of the work rather than a one-time setup.
Developing the visual style
I did not begin with a finished style and then generate the chapters from it. The visual direction developed while I was already working on the first pages.
I wanted the project to take inspiration from both the Attack on Titan manga and anime, while still having an identity of its own. The goal was something slightly more modern and cinematic, but still visibly drawn, imperfect, expressive, and appropriate for the darker atmosphere of the series.
That meant experimenting with things such as:
- line weight and roughness;
- the amount of cross-hatching and screentone;
- contrast between characters and backgrounds;
- how faces and emotional expressions should be rendered;
- how clean or imperfect the linework should feel;
- how much detail should appear in clothing, equipment, Titans, smoke, bone, rubble, and environmental elements;
- how closely the pages should resemble printed manga rather than individual illustrations;
- and how to preserve the atmosphere of Attack on Titan without simply reproducing the exact appearance of either the manga or anime.
I spent several hours refining that direction while the early chapter pages were already being developed. Once I had more completed pages in front of me, I could see problems that were not obvious from isolated test images, so I went back and fine-tuned the style further.
That process is still ongoing.
Some visual characteristics tend to drift during generation. A page may become too clean and digitally polished, too heavily shaded, too noisy, too realistic, too soft, or too generic. Facial proportions, clothing, equipment, line density, and even the apparent drawing technique can also vary from one generation to another.
Because of that, the visual style is not a single prompt that permanently solves the problem. It is a set of evolving decisions, references, constraints, and corrections that have to be applied repeatedly throughout production.
Planning each page
Before generating the artwork, I plan what every page is supposed to accomplish.
That includes:
- the main narrative or emotional purpose of the page;
- the number and arrangement of panels;
- the right-to-left reading order;
- the exact dialogue;
- which character appears in each panel;
- where everyone is positioned;
- what each character is doing or looking at;
- which injuries, clothing details, and equipment must remain consistent;
- the intended rhythm between wide shots, close-ups, action, silence, and dialogue;
- and the final image or line that should push the reader toward the next page.
For action scenes, I also have to keep track of physical geography. Characters cannot randomly change positions, altitude, direction, equipment, injuries, or surroundings between panels simply because a generated image looks dramatic.
For dialogue and Paths scenes, the challenge is different. The page still needs visual tension, readable emotion, and a clear progression, even when little physical movement is happening.
Generating and iterating
Pages are developed panel by panel rather than generated as finished chapters in one attempt.
A large part of the process consists of generating a result, evaluating what is wrong with it, changing the instructions, and trying again.
A panel may be rejected because:
- the composition does not communicate the intended moment;
- the expression is emotionally wrong;
- the character does not resemble the surrounding pages;
- the body language feels generic;
- clothing or equipment is incorrect;
- injuries have disappeared or moved;
- the background breaks spatial continuity;
- the action is difficult to understand;
- the visual style has drifted;
- the panel does not leave enough room for dialogue;
- or it simply does not fit the pacing of the page.
Sometimes a result is mostly usable but one part is wrong. Other times the entire panel has to be replaced. In some cases, several generated elements are combined or rearranged to produce the final composition.
The goal is not just to obtain individually attractive images. The panels need to function together as a manga page and the pages need to remain coherent as a chapter.
Maintaining consistency
Consistency is one of the hardest parts of the project.
AI-generated images naturally vary, even when the instructions and references remain similar. A character's face may subtly change, clothing details may disappear, equipment may be simplified, injuries may move, and the overall rendering style may shift from one panel to another.
The same applies to larger elements such as Titan anatomy, the Founder battlefield, Paths, lighting, contrast, environmental scale, and the amount of texture in the image.
This means continuity has to be checked repeatedly across:
- panels on the same page;
- consecutive pages;
- entire scenes;
- and the chapter as a whole.
I regularly compare new results against previous approved pages and revise the instructions or artwork when something no longer matches.
Even after a visual direction has been established, it cannot simply be left alone. New scenes introduce new problems, and solving those problems sometimes reveals that the broader style guide also needs to be adjusted.
That is why I still refine the project's visual rules as I continue working.
Page assembly and lettering
Once the usable panel artwork exists, I assemble the page myself.
This includes:
- arranging and resizing panels;
- adjusting panel borders and gutters;
- preserving the intended right-to-left reading order;
- cropping or reframing artwork;
- positioning dialogue;
- creating and moving speech balloons;
- adjusting balloon size and spacing;
- making sure balloon tails point to the correct speaker;
- and checking that text does not cover important faces, actions, injuries, or visual information.
Lettering is especially important because generated speech bubbles and text are usually unreliable. The dialogue therefore has to be placed and corrected deliberately so that the page reads naturally.
A panel can contain good artwork and still fail as manga if the eye does not move through it correctly or the dialogue appears in the wrong order.
Manual editing and corrections
I also manually edit the pages in Affinity using a display tablet.
I do not redraw every image from scratch, and I do not want to imply that the project is traditionally illustrated. However, manual corrections are an important part of making the final pages usable and consistent.
These edits can include:
- cleaning visual artifacts;
- correcting clothing;
- repairing straps, uniforms, harnesses, weapons, and other equipment;
- fixing devices or mechanical details;
- redrawing small missing or malformed areas;
- correcting facial details;
- removing duplicated or incorrect elements;
- repairing panel edges;
- moving or reshaping speech balloons;
- adjusting text placement;
- correcting contrast and brightness;
- cleaning random black or white noise;
- reducing muddy shading;
- separating characters more clearly from busy backgrounds;
- and making local changes that would be inefficient or unreliable to solve through another full regeneration.
Manual editing is particularly important because regeneration can fix one issue while introducing several new ones.
For example, regenerating a panel to repair a clothing detail might alter the character's face, pose, expression, equipment, or background. When the image is otherwise usable, correcting the specific problem manually is often the better option.
The manual work and the generation process therefore support each other. Generation provides the underlying visual material, while editing helps preserve the details, consistency, and readability that generation alone does not reliably maintain.
Final cleanup and review
Before a page is considered finished, I review it as part of the surrounding sequence rather than only as a standalone image.
I check:
- whether the panel order is clear;
- whether the dialogue reads correctly;
- whether the characters remain recognizable;
- whether injuries and equipment are consistent;
- whether the environment matches the previous page;
- whether the visual style has drifted;
- whether the important emotional beat is clear;
- whether the page is too dark, too noisy, or too visually flat;
- and whether the final panel creates the intended impact.
The page may then go through additional contrast adjustments, cleanup, text changes, panel replacements, or manual corrections.
Sometimes a page that initially looked finished has to be revisited after later pages expose a continuity or style problem. As the project develops, earlier work can also be updated to better match the refined visual direction.