The Stakes
Gameplay, replays, and the online leaderboard are done — this is the one decision standing between "functional prototype" and "a game with a face." Nothing visual is sunk cost: the current palette, font, and every enemy/boss sprite are stand-ins, not a direction anyone chose on purpose, so this is a genuinely blank canvas rather than a constrained rewrite. The brief below isn't trying to make the call for you — it's trying to make the call fast and grounded in what actually exists today (a working 3D engine, a named roster, a working leaderboard, and a 3D artist who can build real geometry) rather than in assets that don't need protecting — the current placeholder art is explicitly not one of those assets.
Current State
Everything below is placeholder, not commitment — nothing here is a constraint any direction has to honor. What's real is the naming and the leaderboard; the pixels themselves — and the AI-image pipeline that produced them — are fair game, and on their way out regardless of which direction wins, now that real 3D-modeled art is the plan.
- UI Palette
-
Neon OrangeElectric BlueOrange GoldSilver ChromeWhite GlossRed LeatherGreenPlaceholder — up for change along with everything else here.
- Title Font
- Audiowide — currently a stand-in, not a commitment. Any direction is free to replace it; there's no existing identity it needs to stay compatible with.
- Enemy / Boss Roster
-
boss/01boss/02boss/03b boss/04boss/05boss/06b boss/07boss/08boss/09b boss/10boss/11b+ 5 moreThe roster's insect creature names (Weevil Raider, Queen Leviathan, etc.) have since been retired in favor of neutral numeric IDs — see scripts/boss_roster.gd. The underlying three shared art archetypes (warship/dreadnought/leviathan) keep their names below since they're production-pipeline labels, not in-fiction ones.
- Placeholder Enemy Art
- Hard-surface sci-fi PNGs, AI-generated from hand-drawn concept refs, rendered as billboard Sprite3D nodes inside the game's existing 3D engine (idle glow-pulse crossfade). This was a stopgap to get something on screen, not a pipeline worth protecting — every direction below is scoped against real 3D-modeled geometry, not against reusing this AI-image method.
- 3D Production Capacity
- A dedicated 3D artist is now available to model, texture, and rig real geometry for the enemy/boss roster. That's the baseline medium for every direction in this document — there's no cheaper 2D-sprite fallback being weighed against it, and no direction below should be read as if there were.
- Background
- Placeholder only — a scrolling starfield of dots. Open canvas either way.
- Online Leaderboard
- Shipped — replay submission, tag picker, consent flow all working.
- Time Attack
-
Leaderboard category — shipped, not a separate mode. There's only one game mode, and every run always records both score and time locally, win or lose. Submission is manual and deferred — the Replay Browser's SUBMIT button, available for any saved replay regardless of outcome — and a submission always challenges the Score toplist on a valid:true result (best score wins), additionally challenging Time Attack only on victory:true (fastest time wins). Victory requires actually finishing the wave-12 boss's full defeat sequence, not merely reaching wave 12.
Dedicated in-game mode — not built. No timer-focused ruleset, UI, or menu entry exists client-side (checked scripts/) — today it's a ranking lens applied after the fact to the one existing mode, not something a player chooses to play. The chase-the-top-spot hook is still aspirational here; worth designing the visual feedback for now rather than bolting it on later.
Signal Check
What's saturated in shmups/bullet-hells right now versus what's comparatively open — cross-checked against real tag data, not exhaustive.
Candidate Directions
Fourteen case files, A through I in alphabetical order, with J, K, L, M, and N appended at the end. C, D, F, G, H, J, K, L, M, and N are the fully worked options; A, B, E, and I are wildcards worth a look precisely because nobody asked for them yet. Every direction here assumes the same medium — real 3D-modeled geometry, built by the team's 3D artist; the AI-generated 2D placeholder billboards are being retired regardless of which direction wins, so medium isn't a decision left to make. What still varies is production cost: whether a look comes from a single reusable shader or material pass over the roster (most of the set), from bespoke per-asset sculpting and shading (C, E), or from an entirely new rendering pipeline that has to be built before any asset is finished (N). Each card also ends with a copy-paste mood-board prompt (palette hexes and motifs baked in) for Copilot, ChatGPT, Grok, or any similar image generator.
Analog Horror Transmission
You're not seeing the fleet — you're seeing a corrupted broadcast of it. Scanlines, tracking errors, chromatic bleed. Dread delivered by a dying signal, not by creature design.
Bioluminescent Abyssal
Trade disgust for the sublime. The swarm reads as a deep-sea creature that shouldn't be looked at directly — alien and beautiful, not icky. Bioluminescent glow stands in for HP and telegraphs.
Body Horror / Organic
The swarm isn't mechanical — it's alive, and it wants you inside it. Wet chitin splits to show muscle; bosses pulse and breathe between phases. Beauty and disgust in the same frame.
Touchstones — David Cronenberg (The Fly, Videodrome); H.R. Giger's biomechanics without the chrome; Scorn's flesh-and-machine environments; Bugsnax's hybrids played for dread instead of comedy.
Production — The bespoke pulsing/breathing animation is hand-rigged 3D blend-shape work, repeated per creature — that's the real cost driver here, not the modeling itself.
Crystalline Geological
The swarm isn't alive — it's mineral. Geode-like growth splits the hull plating, faceted crystal blooms through cracked metal, and light refracts through translucent growths instead of glinting off armor. Alien and cold, closer to a phenomenon than a creature.
Touchstones — Annihilation's Shimmer and its crystalline mutation set-pieces; Destiny's Vex angular geometry; geode and agate macro photography; the uncanny non-biological hostility of Killzone's alien hardware.
Production — The glow-through-facets shader runs on real modeled geometry, giving true parallax and light-catching as the camera and projectiles move — the facet detail earns its keep specifically because it's sculpted, not painted on.
Ecclesiastical War-Machine
The fleet reads as fallen-angel iconography — gold leaf over the hulls, stained-glass color blocking in the shot effects, baroque ornament where H would put a rivet. Dread delivered as awe, not disgust.
Fungal / Mycelial Bloom
Trade C's wet viscera for drought and decay. Spore clouds drift off cracked, chalky husks; mycelial growth spreads in dry radial fractures instead of veins; the horror is a slow, quiet consumption rather than a wound.
Touchstones — The Last of Us's cordyceps-infected; Annihilation's fungal-bloom sequences; Scorn's drier environmental beats; real bioluminescent-fungus photography for the glow accents.
Production — The dry-decay corruption shader is a material pass over the hard-surface base geometry — reusable across the roster once tuned.
Infested Biomechanical Insectoid
Same war-machines, worse news: something's growing inside the chitin. Hard carapace shells with organic corruption bleeding through the seams — cracked plating, toxin-glow veins, larval growths on otherwise mechanical hulls.
Touchstones — Half-Life's Xen creatures; Dead Space's corrupted-but-readable machines; the underlying art archetypes' own names taken literally — an infested dreadnought, a leviathan mid-metamorphosis.
Production — Hard-surface base geometry plus a single corruption shader on top — the reusable-shader logic the recommendation below is built on.
Retro-Futurist Insectoid
Cold institutional order under threat from an ancient, chitinous swarm. Monolith-slab minimalism and a single warning red, colliding with sand-brown war-machines that look bred, not built.
Touchstones — 2001: A Space Odyssey's HAL-calm and black monoliths; V (1983)'s chrome control rooms and red uniforms; Starship Troopers' carapace texture and propaganda-poster starkness.
Soviet Constructivist Propaganda
Stark red/black/cream, bold diagonal composition, actual agitprop-poster graphic language — applied to the creatures, the UI, and any key art alike, not just the roster.
Lovecraftian Cosmic Horror
The swarm isn't hostile so much as incomprehensible — it barely registers you exist. Silhouettes resolve into shapes that shouldn't parse as ships; the palette drifts into colors that read as faintly wrong rather than dark; killing one doesn't feel like victory, just being ignored slightly less.
Touchstones — Lovecraft's own non-Euclidean-geometry prose; Bloodborne's Great Ones and the Old Hunters' cosmic dread; Color Out of Space's chromatic corruption; Deathstate, the one bullet-hell roguelike already built entirely around this premise.
Production — The geometry-warp shader distorts the 3D mesh's vertices directly — a shader pass over the existing model, not a re-sculpt.
Steampunk Frontier — Space Western
The swarm isn't a hive or a fleet — it's a rustler gang. Riveted brass hulls, exposed piston arms, and gatling-drum weapon pods ride in on oxidized copper trails instead of engine glow; hits crack like revolver fire and leave scorch instead of shatter. War read as a range war, not an invasion.
Touchstones — Firefly/Serenity's browncoat frontier grit and lived-in "used future"; Treasure Planet's sail-and-steam space vessels; Wild Wild West's brass-and-clockwork gadgetry; Sunless Skies' oil-lamp gaslight aesthetic applied to a starfield instead of a sky.
Production — Same hard-surface-base-plus-shader logic as G.
Corporate Dystopia — Hacker-Noir Cyberpunk
Not an arcade cabinet's neon — a surveillance state's chrome. Matte gunmetal hulls stamped with corporate insignia, cybernetic seams glowing cold acid-green through the plating, warning-red telemetry where a face would go. Dread comes from being tracked and owned, not from spectacle.
Touchstones — Blade Runner's rain-and-fog atmosphere used for mood rather than neon signage; Ghost in the Shell's cyberbrain wetware and corporate-military ambiguity; Cyberpunk 2077's Night City corpo-gang factionalism; Deus Ex's augmented-human iconography.
Production — The data-glitch/wetware shader is a material pass over the hard-surface base — same reusable logic as G and K.
Cel-Shaded / Toon-Shaded 3D
Real geometry rendered with thick ink outlines, flat color fills, hard shadow bands instead of a gradient. Reads like a Saturday-morning-cartoon fleet, not a hand-painted one.
Touchstones — Guilty Gear Xrd/Strive, the technical benchmark — its artists hand-edit vertex normals per character to control exactly where shadow bands fall; Hi-Fi Rush's full-world toon shading (detailed enough to get its own GDC talk); Jet Set Radio, the foundational cel-shaded classic; Tunic, cited here only for its fixed elevated camera reading cleanly with 3D models — it's low-poly/flat-shaded, not true ink-outlined cel-shading, so it supports the readability case without being a style match.
Production — Every direction in this document now builds from real 3D-modeled geometry, so modeling itself isn't what sets M apart. The distinguishing cost is the toon-shading pipeline: getting ink-outline thickness and shadow-band placement to read correctly per silhouette usually means hand-tuning vertex normals per model (Guilty Gear Xrd/Strive's own artists do exactly this) — per-asset shader-authoring work that a single reusable material can't fully automate the way G's or K's shaders can.
High-Res 1-Bit Engraving
Running in-engine on regular-enemy tiers 01-04 (assets/shaders/ink_hull.gdshader + ink_outline.gdshader, wired from Enemy._build_model()) — a first real pass on this roster, ahead of the player/boss rollout once more 3D models land. Accent palette below is live: tier 01 cold violet, 02 desaturated blue, 03 acid yellow-green, 04 dusty amber.Primarily 1-bit — pure black and pure white, with every mid-tone built from dense dithering, stippling, cross-hatching, engraving lines, and ordered halftone rather than a flat gray or a smooth gradient. Into that stark field, a restrained set of muted, dirty accent colors surfaces selectively — bioluminescent seams, energy sources, alien fluids, oxidized plating, weapon glow — never a wash of color replacing the monochrome base, always a highlight breaking through the engraving. Each creature carries one dominant accent, or two complementary ones at most, so the roster still reads as a coherent monochrome fleet with just enough color to tell its members apart at a glance. High-resolution low-poly 3D hulls and grotesque biomechanical entities run through a reusable rendering pipeline — geometry → lighting → monochrome conversion → dithering/stippling → engraving/hatching → outlines → selective accent color — that does the finishing work a hand-painted style would otherwise ask an artist to do one asset at a time. The target is modern, high-resolution 3D cosmic horror printed through an impossibly sophisticated 1-bit engraving process — not a pixel-art downsample, not a PS1 filter, not CRT scanlines, and not a simple black-and-white threshold.
Touchstones — World of Horror's 1-bit MacPaint-dithered Junji-Ito cosmic dread; Return of the Obra Dinn's monochrome dithering applied to hard-surface geometry and readable-at-a-glance silhouettes; Mike Mignola's Hellboy/B.P.R.D. inking, where heavy flat blacks carry the whole image and one desaturated spot color does all the atmospheric work; 19th-century steel engraving and stippling illustration; risograph and letterpress print artifacts, whose natural one- or two-ink limitation is the same production logic as this direction's accent palette.
Medium — High-resolution low-poly 3D, and both halves of that phrase matter: "low-poly" keeps geometry economical enough for a small team to produce at bullet-hell volume, while "high-resolution" is what the rendering pipeline needs to make a simple hull read as densely detailed once dithered and hatched. The brief puts the artist's job squarely on strong silhouettes, distinctive alien geometry, and composition — not painting textures or hand-shading every surface. The one addition versus a pure 1-bit pass: an emissive/accent-color channel has to survive the monochrome conversion instead of getting flattened with everything else, so a handful of surfaces per model — a vein, a weapon core, an oxidized seam — need tagging as "color-through" rather than "convert to black-and-white."
Production — Unlike the reusable-shader directions (D, F, G, K, L), N needs a whole rendering pipeline built from scratch — lighting, monochrome conversion, dithering/stippling, engraving/hatching, outlining — before a single asset is finished; modeling the geometry itself is the easy part with a 3D artist on the team. Restrained accent color adds one more small step on top: picking one or two hues per creature from a shared palette and tagging which surfaces carry them — a per-asset decision, but a cheap one, a color choice and a mask rather than new geometry or a new shading pass. The payoff is still a single reusable pipeline rather than a per-model shading job — tune it once and it runs over the entire roster, environments, and UI alike, closer to I's "reskins everything for free" leverage than to M's per-model outline work. The cost sits upfront and entirely on the tech side: building a dithering/stippling/lighting/outline pipeline sophisticated enough that a relatively simple low-poly model can acquire a highly detailed appearance without hand-painted shading — get it wrong and it reads as a cheap black-and-white filter, not an engraving. Once that pipeline is mature, new assets become a silhouette-and-geometry-and-one-accent-color problem rather than a texturing problem, which is exactly where a small team's — and their 3D artist's — time is best spent.
Nothing above requires picking exactly one letter. G is proof this document already mixes directions on purpose — its hard-surface base is H, its corruption is C, blended deliberately rather than chosen between. Read the rest of the set the same way: as a parts bin, not a ballot.
The cheap mixes are the ones that stay inside the same production logic — a handful of these directions (D, F, G, K, L) are each built around one reusable shader over a shared hard-surface base, so swapping which shader runs, or layering two of them across different roster tiers, costs little more than picking one. C and E break that logic — their cost is bespoke per-asset work, so folding either into a mix reintroduces C's full price tag rather than sharing it. A's screen-space signal-corruption and I's flat shape-and-type both sit a layer above everything else (they're presentation filters, not creature material), which is exactly why they're the two easiest to bolt onto any other direction rather than merge with one. M and N break the logic in a similar way to each other: neither is expensive per asset once built, but each is a global final-render treatment applied to the whole image — cel/toon shading for M, the lighting→monochrome→dither→engrave→outline pipeline for N — rather than a material layered onto a base the way G's corruption shader or K's patina are. You can put G's toxin-glow texture on a model destined for either treatment, but you can't run M's toon shading and N's 1-bit engraving on the same frame at once: picking one commits the entire visual language, not just the creature material, which is why neither slots into the cheap, blendable cluster above.
Tone can clash even when cost doesn't: L's quiet surveillance-dread and K's loud gunslinger energy are both cheap, shader-driven directions, but stapling them together undercuts both rather than combining them. The mixes worth prototyping first are the ones where two directions answer different questions — G for "what are these enemies made of," B for "how do bullet patterns read," I for "how does the UI look" — rather than two directions competing to answer the same one.
Decision Matrix
Score these yourself — this is a starting point, not a verdict. Dots indicate magnitude of the row, not good or bad.
| A · Analog Horror | B · Abyssal | C · Body Horror | D · Crystalline | E · Ecclesiastical | F · Mycelial | G · Infested Hybrid | |
|---|---|---|---|---|---|---|---|
| Art complexity per asset | |||||||
| Market differentiation | |||||||
| Reusable VFX leverage | |||||||
| Readability at speed | |||||||
| One-line read | Almost entirely a screen-space effect — cheapest total rework, works over whatever creature art gets chosen underneath. | Glow/bloom system reused across the roster, and it doubles as gameplay clarity. Underrated for the cost. | Most distinctive per asset, but bespoke animation work doesn't scale across 16+ enemies — the expensive direction, full stop. | Cold and alien without being gross — a third material logic (not flesh, not metal) that turns bullet-hell readability into a built-in feature via glow-through-facets. | Gorgeous prestige-horror via reverence, but ornate per-asset detail (gold leaf, stained glass) is closer to C's bespoke cost than a shader — the priciest wildcard here, and busiest at speed. | Same organic dread as C, dry instead of wet — cheaper VFX (particle dust vs. wet shader) and a texture language nothing else here is using. | One strong reusable corruption shader carries most of the roster; hard-surface base is a proven technique. Efficient way to earn the horror. |
| H · Retro-Futurist | I · Constructivist | J · Lovecraftian | K · Space Western | L · Cyberpunk Noir | M · Cel-Shaded 3D | N · 1-Bit Engraving | |
|---|---|---|---|---|---|---|---|
| Art complexity per asset | |||||||
| Market differentiation | |||||||
| Reusable VFX leverage | |||||||
| Readability at speed | |||||||
| One-line read | Cheapest to execute cleanly since it needs the least visual reinvention from what's already there. Risk: closer to "competent sci-fi" than "unforgettable." | Cheapest and most differentiated entry on this list — flat shape and type instead of rendered art, and the only direction that reskins UI and marketing for free. | A reality-warp shader carries the roster efficiently, but "wrong geometry" cuts against the clean silhouettes bullet-hell readability needs — same tension as E's ornament, different cause. | Lowest signal-check score on the list — an uncontested mashup. Same reusable-shader logic as G, tone doing the differentiation instead of body count. | Sidesteps the crowded Tron-grid/synthwave lane for surveillance-state dread. A data-glitch/wetware shader carries the roster efficiently, but "cyberpunk" itself caps the differentiation ceiling. | Strongest whitespace claim on the list — near-zero genre precedent — but the outline/toon-shading pipeline has to be built and tuned per model, not reusable like a single material shader. No track record on bullet-hell readability. | Needs the same from-scratch pipeline investment as M, just a different one (dither/engraving, not toon-outline) — but it's a single reusable pass, cheaper to extend than M's per-model work. Zero bullet-hell precedent for 1-bit engraving. |
A Lean, Not a Verdict
With nothing visual actually locked in, every direction here starts from the same blank canvas — so the honest lens isn't "what protects existing work" but "what's cheapest to make look good across 16+ enemies, and what genuinely escapes the pixel-art crowd." On that basis, G — Infested Biomechanical Insectoid still comes out ahead, but for a cleaner reason than before: a hard-surface base is cheap to build once as real 3D-modeled geometry, and a single reusable "infection" shader (cracked plating, toxin-glow veins, pulsing corruption) can then be layered across the whole roster instead of hand-crafted per creature. It also honors the organic/horror pull that's been growing on you without paying C's full bespoke-animation cost, and it leaves room to borrow B's best trick later — toxin-glow doubling as a bullet-pattern telegraph, a gameplay win, not just a mood one.
Set against the two propaganda-adjacent wildcards, G holds up for a different reason than it beats C. E — Ecclesiastical War-Machine is the most emotionally distinct option in the whole set — reverence instead of disgust, a register nothing else here reaches for — but its gold-leaf-and-stained-glass ornament is per-asset work in the same way C's bespoke animation is: it undercuts the "one reusable shader" logic that makes G efficient, and it scores worst on readability at speed of anything in the matrix, a real risk in a bullet-hell. I — Soviet Constructivist Propaganda actually beats G on pure efficiency: flat shape-and-type is the cheapest asset class scored here, it's the single highest-leverage system in the document because it reskins UI and marketing for free alongside the roster, and it posts the best readability score in the matrix. The catch is scope, not cost — I isn't a creature reskin, it's a full identity change; picking it commits the menus, the leaderboard screen, and any key art in the same motion as the roster, and a poster-flat swarm doesn't answer "what is attacking me and why" the way G's infested-hardware premise already does with zero new lore. G stays the lean because it's the only one of the three that upgrades the current placeholder without re-authoring everything around it — but I is the one worth stealing from if the team decides the game's whole presentation, not just the enemy roster, is in play.
C is the strongest single idea in this document but the most expensive — worth keeping alive as an "if G feels too safe once mocked up" escalation rather than the default. A and B are worth taking seriously on their own merits now, not just as consolation prizes: both lean almost entirely on one reusable system (a broadcast-corruption shader for A, a glow/bloom system for B) rather than per-asset art, which makes them arguably cheaper than G to prove out — and B's glow language solves a real bullet-hell readability problem regardless of which direction wins.
Concrete next step: pick one boss — boss/12f, the final-boss slot still wearing the ascended_queen_leviathan archetype art, is a good test case, since that art already implies a metamorphosis — and do a single concept pass in two or three of these directions before locking the roster-wide decision. That's a day of art, not a month, and it turns this whole document from a set of opinions into something you can actually look at.