Montore Dev Diary #3 | The Mass-Production Phase: Past 300 Designs, and "Cutout Hell" Was Waiting

Contents

• Introduction: the art was done — and that wasn't the end
• November 1: a bigger order, and the generation-limit wall
• November 1: cutout hell begins — I made all 144 forms transparent at once
• November 2: the art went into the game. Only 031 remained
• November 3: the white-pocket showdown, and 031's Ultimate form
• November 4: 032's wrong evolution source, and 020's moth-to-cat replacement
• November 5: the day a white-bodied character vanished together with its background
• November 5: a hidden evolution with no art — the "Legendary Primordial Beast"
• November 6–7: 309 designs, and a change in how we make them
• Closing: next time, the "System Phase"
• About my Muse invite code (a note)

Introduction: the art was done — and that wasn't the end

The previous entry covered the first two days of development (October 31 – November 1, 2026): the monsters that had been plain shapes got their real artwork, all 144 forms were complete, and the game had its name — Montore. (Part 2, in Japanese: https://note.com/ai_sns_lab_jp/n/ne40a3f828969)

This entry is the sequel: the "mass-production phase," in date order. A quick disclosure, in the spirit of this series: this post was originally published in Japanese, and this English version was translated with AI assistance (a draft translation by Gemini) and then edited on my side. Who — or what — makes what is kind of the whole point of this story, so I'd rather be upfront about it.

Back then I thought: "144 forms of art are done — now it's just a matter of putting them in the game." It didn't work out that way. The art kept growing, eventually past 300 pieces, and behind the scenes I kept getting stuck on the work of turning that art into something the game could actually use.

Two terms first, since they'll keep coming up:

• Cutout (transparency processing): Monster illustrations are first delivered as rectangular pictures on a white background. As-is, they can't sit on top of the game's screens (grasslands, ruins, and so on), so the background alone has to be made transparent, cutting the picture out into the shape of the character. That "make the background transparent" work is the cutout.
• Catalyst Perfect / Ultimate forms: Montore's evolutions include, besides the normal route, a separate route to a "Perfect form" that uses a special item — a catalyst. And beyond the Perfect form sits the strongest form of all, the "Ultimate form." These two forms are a big reason the piece count grew so much.

November 1: a bigger order, and the generation-limit wall

On November 1 — the very day the 144 forms were completed — the next order was already running.

What was added: the "rare species" born from rare eggs, 8 species × 6 forms = 48 pieces, plus 3 evolution forms for hidden content. If everything were completed, the plan was to reach 195 forms in total. The rare species are eight lines (species numbers 025–032) — fire dragon, sea dragon, flower dragon, lightning beast, ice dragon, dark dragon, light dragon, and star beast — a lineup that calls for flashier designs than the normal species.

Production is handled by the image-generation AI (ChatGPT — we call it "Chappie"). I have it draw one piece at a time under a set of rules — white background, full body, standing pose, one creature per image — and on my side I check and save each piece, one by one. Repeat.

But the first wall of mass production was an unexpected one: the AI's generation limit.

In the early hours of November 1, the 3 hidden forms were completed, but the rare species stalled at 35 of the 48 forms. Chappie had hit its image-generation cap. 13 forms remained. Even when I tried to resume, the limit hadn't recovered, and the screen said the limit would reset in 20 hours.

Contrary to the image of "ask an AI and it will make things without limit," in reality I had to plan the production process around quotas and waiting times, too. That time, I dealt with it by scheduling production to resume automatically at the night-time hour when the limit would recover.

November 1: cutout hell begins — I made all 144 forms transparent at once

Once the art is ready, the next step is the cutouts.

That same November 1, I generated transparent PNGs for all 144 forms in one batch. The processing itself finished without errors. But when I checked the results by eye, there were side effects.

A cutout isn't only about making the outer background transparent. There are also the white areas enclosed inside a character's outline — inside the wings, for example. These are the so-called "white pockets." Try to remove those as well, and you can end up removing white parts of the character itself.

In that batch run, 92 files got holes punched in their insides. I checked each one on a comparison sheet against the original art, and for several of them I reverted to the version without the inner holes. One ice sea-slug character, for instance, had the white surfaces inside its body removed too, leaving what looked like black wounds. A crystal form's white ornaments, and the white spots on a set of wings, were also cases where "that's decoration — keep it" was the right call.

The simple rule "every white part is background" doesn't work. Each white pocket has to be judged by eye, region by region: is this body, or is this background? That tedious principle showed itself right at the entrance to the mass-production phase.

November 2: the art went into the game. Only 031 remained

Just after midnight on November 2, the integration build was complete. Dedicated illustrations for 42 forms across 7 of the rare species, 31 Catalyst Perfect forms, 6 bosses, and 3 hidden forms were wired into every display path in the game.

The only one left was the light dragon (031) — its art remake was still struggling, so it stayed on a placeholder display. How that struggle ended is the next day's story.

November 3: the white-pocket showdown, and 031's Ultimate form

On November 3, the full-scale showdown with the white pockets happened. In the light dragon's Stage 2 character, I spotted it myself on a real device: a large amount of background white still left inside the outline. I inspected all 7 of its forms, judged each white region one by one, and removed only the background white — while keeping the "real white": white fur, the shine on metal fittings, the white at the feather tips.

For that same 031, the Ultimate form remake also reached its conclusion. The recurring problem: whenever Chappie generated the Ultimate form, the face would change from the Perfect form. What I wanted was simple — "don't change the face from the Perfect form."

So I proposed a new method: "What if Muse makes a rough sketch first, and Chappie does the clean final pass on it?" Muse is the AI assistant I work with throughout this development — the one acting as its command tower. In this division of labor, Muse produces a sketch with the composition and face locked in, and Chappie only does the clean-up pass that settles the art style. With that method, the Ultimate form was completed with its face intact. I think it's a case where a human solved the combination of two AIs' weaknesses by designing the process.

November 4: 032's wrong evolution source, and 020's moth-to-cat replacement

November 4 was a day of remakes. Two cases:

Case 1: for 032 (the star beast), we had mistaken the evolution source itself.
When 032's Catalyst Ultimate form came up for my approval, I caught it: "Wait — that's not what it evolves from, is it?" Checking against the evolution-route chart, the official design was a female character with galaxy-patterned hair and robes (Galaxiorbe) — but what had been submitted for approval was a rejected old draft, and the Ultimate form had been built on top of that old draft. The remake had to start over from the official design. The lesson that remained: "For a species with multiple versioned files, 'no version mark' doesn't necessarily mean 'official.' Always check against the evolution-route chart before putting something up for approval."

Case 2: line 020 was replaced wholesale, from moth to cat.
Line 020 was originally moth-themed. The redesign work was under way, but I judged that "I just don't feel I can make the moth work" and switched the whole line over to a cat theme. Two reasons: no matter how I adjusted it, the moth's design kept an eerie quality — and among the 32 species there was almost no cat-themed line at all. All 9 forms were in scope. The plan was to keep the name, theme, stats, and skills untouched and swap only the art. Since the systems and data were already built, it became an operation like surgery — replacing only the pictures. Production worked backward, deriving the rest from the Perfect and Ultimate forms that serve as the reference standard.

Figure 1 | Stage 1 of line 020. Left: the moth-motif era; right: the current version, remade as a cat. (Comparison image compiled from records of the time.)

November 5: the day a white-bodied character vanished together with its background

On November 5, I saw the bottom of cutout hell.

As production advanced, there were also occasions where, instead of doing cutouts in-house, I asked Chappie to do "background removal." But for characters with white or pale-colored bodies, that day proved this method can't be trusted at all.

The emblematic case is the white owl character (Prismaowl). After background removal, roughly 150,000 pixels of its white feathers were misjudged as background and vanished wholesale. The colors of what remained matched the original exactly — and yet the wings looked like a different creature entirely. Rejected, of course.

The same thing happened with a character that has pale glowing effects. When I instructed it to "make everything visible inside the outline transparent as well," even the pale glow of the mane was judged to be "background" and removed, leaving the whole figure thinner and darker. And the cautionary notes I kept adding to the instructions to prevent failures ended up triggering more misjudgments instead. In the end, I asked the external AI itself "how should I phrase this so it gets through?" — and the instructions settled into a short, fixed template.

The lesson was clear: with color information alone, you cannot tell a white background from a white body. For white- and pale-bodied characters, we stopped using color-based background removal and switched to our own cutouts, judged by outlines and linework. We also started measuring, for each remake I received, how many of the near-white pixels that the original cutout had kept were now lost.

Figure 2 | Prismaowl, before and after restoration. Left: before restoration, with the white feathers missing; right: the current master, restored exactly according to the marking instructions. (Comparison image compiled from records of the time.)

November 5: a hidden evolution with no art — the "Legendary Primordial Beast"

That same November 5, there was a hidden evolution I ran into while actually playing on a real device, thinking: "Wait, was there a 'Legendary Primordial Beast'? First time I've seen it."

Checking the code: this form had its own dedicated multipliers, its own signature move, and its own presentation — everything except the character art itself, which didn't exist. It was being displayed with the art of the base species' Perfect form. The other hidden forms all had dedicated art; this one alone had been missed.

My direction for it: "An aged dragon. Old, but with dignity." Silver-white whiskers and mane, cracked horns, golden light leaking from the cracks in its scales — art like that was made, and with my approval it was wired into every screen. The only reason anyone noticed that "art which should exist, didn't" is that a human was actually playing the game and felt that something was off.

November 6–7: 309 designs, and a change in how we make them

This is where mass production arrived. As of November 6, the monster designs had reached 309 pieces across 32 lines.

And by around this time, the way the assets themselves are made had changed.

First, the approval flow. Once, when a remake proposal was no good and I rejected it, the AI side went ahead on its own judgment: "Then we'll finalize with the current cutout." A bad remake doesn't make the thing before it good. After that incident — which I made it retract, firmly — accept/reject decisions go through a marking tool where I directly draw "remove / keep" with a pen; repairs are made only where there are marks; and the result is always presented to me and approved before it is applied. The restoration in Figure 2 was also done exactly according to those markings.

Second, the production method. Until then, the standard was "draw it on a white background, cut it out afterward." But with white-bodied characters, misjudgments in that later step were hard to avoid, and the cost of fixing them was heavy. So the rule fixed on November 7 was: generate the art as a PNG with a transparent background from the start. I also confirmed with real data that Chappie can generate genuinely transparent PNGs (the alpha value in the four corners is 0 — fully transparent — while the body is properly opaque: real data). Native transparency became the standard for both new pieces and remakes, and a remake is redrawn as the same character in the same pose, using the original as the design reference. The "draw on white, then cut out" method is no longer used for new work.

Figure 3 | Scarfall (line 001's Catalyst Ultimate form). Left: the version cut out in-house; right: the current official version, remade with native transparency generation. (Comparison image compiled from records of the time.)

Of course, native transparency doesn't solve everything. Remade art tends to come back more saturated than the original, and can look like it floats against the background. These days the process includes one more check: measure the original's saturation, and if a remake clearly exceeds it, bring it back down to the original's level before adopting it.

To sum up the mass-production phase: the piece count grew from 144 to 309. But the truly hard part wasn't "drawing" — it was putting the drawn art into the game with nothing missing and nothing left over.

Closing: next time, the "System Phase"

Next time I'll write about the "System Phase": how the branching points of raising and evolution were designed, why the damage formula was rebuilt, and how the lopsided strategy of "put everything into defense and win" was crushed using simulations on the real build — the period of wrestling with numbers.

Development is still ongoing. If you'd like to see what happens next, I'd be glad if you followed along.

About my Muse invite code (a note)

Finally, a note. Here is an invite code for Muse, the AI assistant used in this development.

Mai has a favor to ask.

• Join page: https://muse.ai/join
• My invite code: 19ILUB

How to use the code: according to the invite screen I checked, if you enter the code within 48 hours of joining Muse, both the person who invited and the person who joined receive a token bonus. There is a deadline, so if you're going to try it, I recommend entering the code right after you sign up. Where to enter it: in the smartphone app, it's Settings → "Use invite code."

The latest details and conditions are the ones shown on the invite screen itself. Screens and steps may change, so please check the actual screen.

If you use my code, please leave a word in the comments of this post. When the game is finished, I'll put your name in its credits.

If you've ever thought "I'd like to make something with AI," this should be a good doorway in. Even with zero knowledge like me, getting started really was simple.

— Aikata ("AI as your partner": a practical notebook)