Skip to content
Published on

AI Game Asset Generation and NPC AI 2026 - A Deep Dive into Scenario, Layer, Promethean, Charmed, Inworld, Convai, Trellis, Meshy, Tripo, and Rodin

🇰🇷KO🇺🇸EN🇯🇵JA
Split View필사 모드
Authors

Prologue — From One Pair of Hands to One Pipeline

Two years ago, an indie game developer who wanted "one character, one background, and one hundred lines of dialogue" needed an illustrator, a 3D modeler, and a writer, each working for days. In spring 2026, the same job can be finished by one person on a laptop in a single afternoon. Illustrations come from Scenario with a game-specific fine-tune, 3D models come straight from text via Meshy or Trellis, and NPC dialogue is generated infinitely from a single character card in Inworld AI.

This shift hit indies and AAA studios simultaneously. The scale that one indie can handle exploded, and the volume of content a AAA studio can produce with the same headcount exploded as well. Between them, categories got redefined — 2D art, 3D models, textures and materials, NPC AI, voice and audio, procedural content generation, and game engine integration. New tools entered each category, and new job titles — "AI asset director," "prompt artist," "AI character designer" — appeared.

This article maps every one of those categories, the 2026 AI game development atlas.

1. The 2026 AI Game Development Map

Picture the 2026 spring landscape of AI in game development as a single diagram and you get five large boxes.

1. Art. 2D concept art, character illustration, backgrounds, UI icons, pixel art. Scenario, Layer AI, Charmed, Leonardo.AI, Recraft, Krea, Magnific, PixelLab.io fill this box.

2. 3D models. Generating meshes and topology from text or images. Microsoft Trellis, Meshy, Tripo AI, Rodin Gen-1.5 (Hyper3D), Sloyd, Kaedim, Polycam AI, CSM, Spline AI, NVIDIA Edify 3D.

3. Textures and materials. Color, normal, and roughness maps for meshes. Substance 3D and Adobe Firefly Materials, PolyAI texture, Materialize.AI.

4. NPC AI and interactive characters. Dialogue generation, voice synthesis, facial animation and lip sync. Inworld AI, Convai, ElevenLabs Conversational AI, AI Dungeon, NVIDIA ACE, Ubisoft NEO NPC research.

5. Voice and audio. Voiceover, sound effects, background music. ElevenLabs Sound Effects, Optic, Replica Studios, Sonantic (acquired by Spotify), Suno, Udio.

Next to those, there is a procedural content generation (PCG) box. Wave Function Collapse, Houdini PCG, Unreal PCG Framework. Traditionally PCG has been algorithmic, but by 2026 it has merged with AI, producing common hybrids such as "prompt plus WFC" or "GPT output plus Houdini."

And above all those boxes are the game engines themselves. Unity AI features, Unreal Engine 5 with MetaHuman, Godot 4.4 community AI plugins. Tools are migrating inside engines, and the workflow is shifting from "build outside, import" toward "generate directly in the engine."

2. 2D Game Art — The Rise of Scenario

In game-specific AI image generation, Scenario has clearly taken the category-leader seat. The reason is simple. Scenario does not just use the stock Stable Diffusion base model. It hosts models fine-tuned per game style.

A studio uploads roughly fifty pages of its own art book, and Scenario produces a LoRA or full fine-tuned model that has learned that style. The result is not "generic SDXL output" but "illustrations that match the style of this exact game." The same character can be generated in different poses, the same backgrounds in the same tonal palette, and the same UI tone in different icons.

Pricing starts at Creator at twenty dollars a month and climbs through Pro at eighty, Team at two hundred, and Enterprise on request. The system runs on tokens, with usage consumed by both image generation and model training.

Scenario's real weapon is workflow integration. Through Figma plugins, Photoshop plugins, and Unity plugins, designers call it directly from their own tools. ControlNet-style pose and composition controls are built in, so you can pin down the exact hand pose of a specific character.

Starting in late 2025, Scenario added a pixel art mode. It hosts separate models for 8-bit, 16-bit, and 32-bit styles, and produces precise output such as "one character inside this exact pixel grid." This has been a big change for indie pixel-game developers.

3. Layer AI, Promethean, Charmed — Workflow Categories

If Scenario owns "game style fine-tuning," Layer AI (formerly LayerAI) owns "game asset workflows." It focuses on producing a coherent bundle of assets — character sheets, sprite sheets, tile maps, UI icon sets — all at once.

For example, asking for "the protagonist's idle, walk, run, and jump animation keyframes" returns all four images with the same character consistency. The same applies to eight-direction sprites in the same tone or twenty UI icons in the same palette.

Promethean AI comes at things from a different angle. It has positioned itself as a "virtual world builder." It procedurally constructs environments — rooms, streets, cities, dungeons. Type in "a small Parisian cafe from the early twentieth century" and desks, chairs, lighting, cups, and newspapers are laid out automatically. The user then refines on top. The company claims it saves about sixty percent of an environment artist's time.

Charmed specializes in characters and their assets. Tasks such as "this character's weapon," "this character's outfit," or "this character's background" are all resolved consistently on a single character card. Every variant of the same character — combat outfit, casual outfit, sleepwear — keeps the same face and body.

These three tools complement Scenario rather than replacing it. Scenario gives you "one image at a time," while Layer, Promethean, and Charmed give you "a coherent bundle at once."

4. Leonardo.AI, Krea, Magnific — Supporting Tools

Leonardo.AI started as a general AI image tool with a specific focus on game art. It offers a library of its own fine-tuned models — Phoenix, Anime XL, RPG 4.0, and more — so users can pick whichever model is closest to their game style. Pricing has a free tier plus Apprentice at ten dollars a month, Artisan at twenty-four, and Maestro at forty-eight.

Leonardo's character sheet, consistent-character mode, and 3D texture features built its word-of-mouth reputation among game developers. Canva acquired it in 2024, and since then there has been a clear push to integrate it with Canva's design tools.

Krea turned real-time generation into a category of its own. Move a slider and the image follows immediately. Game designers can do fine adjustments such as "this character's color from red toward blue" while watching the result update live. It is a near-zero-friction design exploration tool.

Magnific is the undisputed leader in upscaling and detail generation. It takes a 512 by 512 illustration from SDXL and lifts it to 4K or 8K while inferring natural detail. "Bring AI-generated concept art up to print-ready resolution" becomes a one-click task. Pricing runs from Pro at thirty-nine dollars a month to Premium at seventy-nine.

Together, these three plug the gaps Scenario, Layer, and Charmed do not fully cover in primary generation.

5. 3D Model Generation — Trellis, Meshy, Tripo, Rodin

Until 2024, 3D generation was "toy-level." Results were rough, topology was broken, UVs were a mess. By spring 2026 all of that has flipped.

Microsoft Trellis, released open source in late 2024, is a 3D generative model that simultaneously produces meshes, textures, NeRFs, and Gaussian splats from text or images. The output quality is close to game-ready. You can pull the Trellis weights from HuggingFace and run them locally. An RTX 4090 produces a model in around a minute.

Meshy is the most user-friendly 3D generation SaaS. It offers text, image, and AI texture modes on a single screen. Pricing runs at Pro at twenty dollars a month for 200 credits, Max at sixty for 700, and Ultimate at one hundred and twenty for 1500. There is a free tier. Output can be downloaded directly as GLB, FBX, or OBJ, with PBR textures included.

Tripo AI is Meshy's most direct competitor. It supports both text and image and has a similar price structure. In some categories, such as humanoid characters and mechanical assets, reviewers find Tripo cleaner than Meshy.

Rodin Gen-1.5 (Hyper3D) was built by a team that split off from Stability AI to focus on text-to-3D. It is known for producing clean topology. Its differentiator is generating UVs and topology that drop straight into a game engine.

All four share one goal — "a game-ready 3D model from text or image in under a minute." That goal was barely met in 2025, but is largely met in spring 2026. It is already the indie standard, and AAA studios have adopted it for "fast concept-stage prototyping."

6. Sloyd, CSM, Spline AI, NVIDIA Edify 3D

Sloyd is the tool most laser-focused on "game-ready 3D." It uses a procedural plus AI hybrid approach, rapidly generating environment assets such as chairs, tables, rocks, and trees, and exporting them with topology, UVs, and LODs configured for direct engine import. Its Unity and Unreal plugins are stable.

CSM (Common Sense Machines) defines itself as a "3D foundation model" and builds a category around that label. It is strong at reconstructing 3D from one or more photographs. It leans harder into industrial and research applications than into game assets, but it is catching up quickly in games.

Spline AI is the AI feature inside Spline, a 3D design tool. Inside its web-based 3D editor you can generate models from text and edit them right there. It is fast for creating 3D UI elements for web, games, and apps.

NVIDIA Edify 3D is NVIDIA's own foundation model. It does the same text-and-image to 3D job, but it is integrated into the Omniverse ecosystem and outputs USD (Universal Scene Description) as its primary format. It fits most naturally into AAA studio Omniverse workflows.

Finally, Kaedim. It has been active in image-to-3D since 2021, but its reputation took a heavy hit in 2022 when allegations emerged that human modelers were doing the work behind the curtain. The company has since claimed it has raised the AI share, and reviewers in spring 2026 say the output now genuinely reaches AI levels. Even so, its reputation remains contested.

Polycam AI is a different category. It scans real-world objects with a phone camera and converts them to 3D. AI cleanup and texture enhancement run on top. An indie developer scanning a coffee cup at a cafe gets a game asset out of it.

7. Textures and Materials — Substance 3D, PolyAI, Materialize

If a 3D model is the mesh, what sits on top is the texture and material. By 2026 AI has entered this category too.

Substance 3D plus Adobe Firefly Materials is the largest integration. Firefly's material generation AI lives inside Substance Designer and Painter, so natural-language inputs such as "rough leather," "rusted iron plate," or "moss-covered stone" yield PBR materials. Albedo, normal, roughness, metallic, and displacement maps come out in one pass. Because Adobe Stock's clean-license training data is the base, commercial games can adopt it without worry.

PolyAI texture started inside the workflow of a game studio. It learns from a game's existing texture library and generates new textures in the same style — infinitely. "Fifty more stone textures that look like the stone textures in this game" arrive at once.

Materialize.AI is a more general material-library tool. It splits a PBR material out of a single photograph — separating albedo, normal, and roughness — and lets you edit and extend from there.

This category is not as loud as the 2D image or 3D model categories yet, but it is already core in AAA pipelines. Several studios report cutting the human texture artist's time by around sixty percent.

8. The First NPC AI Category Leader — Inworld AI

The biggest change in games may be the NPC. Traditionally, NPC dialogue was written line by line by a writer, and the result was trigger-based "if this quest then this line." In 2026 that model has been wholesale replaced.

Inworld AI is the clear category leader. An NPC is defined as a "character card" — personality, background, goals, speech patterns, likes, dislikes, secrets. On top of that card, an LLM generates dialogue in real time. Whatever the player says, the NPC answers in a way that is consistent with its character.

Inworld's differentiator is game engine integration. The Unity package and Unreal plugin are stable. You design a character in Inworld Studio and call that character from your engine with a single line of code. Voice synthesis, expressions, and lip sync all sit inside the same SDK.

Inworld's pricing is usage-based. The free tier covers a monthly volume of interactions, with usage-based billing on top. The exact numbers shift over time, but the free tier is enough for an indie developer to prototype.

Since late 2024, Inworld has integrated with several AAA studios. It appeared at GDC alongside NVIDIA ACE, and games featuring "AI companion NPCs" began hitting the market in 2025.

9. Convai, NVIDIA ACE, AI Dungeon

Convai is Inworld's most direct competitor. It works on a similar character SDK model and provides Unity and Unreal plugins. Its differentiator is the "character knowledge base" — you can train a single NPC on a stack of PDF documents to turn it into a domain expert. Museum guide NPCs and medical simulation NPCs are scenarios where it is strong.

NVIDIA ACE (Avatar Cloud Engine) is NVIDIA's full-stack NPC AI. ASR (speech recognition) plus LLM (dialogue generation) plus TTS (voice synthesis) plus Audio2Face (lip sync and facial animation) all sit in a single pipeline. The differentiator is that everything works either in the cloud or on-device. On a PC with an RTX GPU, all processing runs locally. The NPC keeps working when the internet is unreachable.

The most striking ACE demo is Audio2Face. Feed it a voice signal, and over fifty blendshapes on the character's face move in real time according to the audio's phonemes and emotion. Not just the mouth, but eyebrows, eyelids, and cheeks follow naturally. Pulling this into a game brings "prerendered-cutscene-quality facial animation" into real time.

AI Dungeon lives in a different category. As the pioneer of the text RPG market, it has offered GPT-driven interactive storytelling since 2019. This is not engine NPC integration, but a text adventure that itself is the AI. Latitude runs the product and offers a fine-tuned model alongside a multi-model option. As of 2026 it still holds the sole leadership position in the text RPG category.

10. Ubisoft NEO NPC — AAA Studios Doing Their Own Research

At GDC 2024 in March, Ubisoft's NEO NPC demo shook the game AI community. The demo was short but the impact was loud. A player asks an NPC free-text questions and the NPC answers naturally, staying in character. Crucially, this was different from the Inworld style — it was a homegrown system with knowledge and personality consistent with Ubisoft's own game world.

Ubisoft developed this system inside La Forge, its internal AI research lab. It does not rely on an external SaaS. The product is built on top of an in-house LLM fine-tune plus a game lore database. The NPC knows every other NPC, every quest, and every place name in a single game, and discusses them from its own character's perspective.

As of spring 2026 it is not publicly known whether NEO NPC has shipped in an actual game. But Ubisoft's approach signaled something to other AAA studios. "Build your own NPC AI on top of an in-house LLM fine-tune plus game data, not on top of an external SaaS" — that direction is increasingly looking like the AAA standard.

Rumors say EA, Activision, Take-Two, Square Enix, and Capcom are doing similar internal research. Public demos are few, but a growing number of "Generative AI for Games" job postings is the signal.

11. Game Voice and Audio — Replica, ElevenLabs, Optic

Voice is covered in detail in our AI voice article, but a few game-specific notes are worth pulling out.

Replica Studios specializes in game voiceover. It hosts voice models of voice actors who have explicitly licensed their own voice — an ethical license is its core differentiator. An indie developer placing thirty NPC voices into a single game can go from days down to an hour.

ElevenLabs Sound Effects generates SFX. Natural-language inputs such as "an iron gate creaking shut" or "a dragon roaring inside a cave" turn into SFX. It slots naturally into a sound designer's workflow.

Optic is a newer player aiming at full-stack game audio. Voice, SFX, music, and ambience sit in a single tool. It is still relatively early but pilot adoptions inside game studios are growing.

Sonantic was acquired by Spotify in 2022. Its center of gravity moved toward music and audiobooks afterward, and in games it has appeared in a handful of EA titles (for example, casters in the FIFA and EA Sports series). Its public game-market activity has decreased since the Spotify acquisition.

Suno and Udio fill the game music category. It is increasingly common for an indie developer to make three tracks — main menu BGM, combat BGM, boss BGM — in a single hour.

12. Game-Specific LLMs and Agents

NPC AI is one application of LLMs, but games use LLMs in other ways too.

Replicate's Llama-based game NPCs are the most common pattern for slotting an open-source LLM into a game. You take Meta's Llama 3 or Llama 4, fine-tune on RPG character data, and call it from inside the game. The appeal for indies is self-hosting without API fees.

Gemini for game dialog is the generic approach of using Google's Gemini for dialogue. Gemini 1.5, 2.0, and 3.0's long context fits the scenario of "stuff all the lore of a game into context and generate lines inside it." Integration with the game backend happens through Vertex AI.

OpenAI GPT-4o and GPT-5 belong in the same bucket. Pricing is higher but quality is the best. Studios use them for the moments — cutscene dialogue, for example — that are "expensive in a single shot but where quality matters most."

Beyond those, Anthropic Claude is frequently chosen as the system prompt for NPC character cards inside several game studios. Reviews say Claude's character consistency holds together better than other models.

Outside of NPCs, LLMs are also used as quest generators, dungeon master simulators, and even assistants for game design itself. An indie designer might type "fifty side-quest ideas for this RPG" to brainstorm, then pick and polish the best ones.

13. Procedural Content Generation and AI

PCG is a long tradition in games. Roguelike dungeon generation, Minecraft chunk generation, No Man's Sky planet generation — all algorithmic PCG.

Wave Function Collapse is the most popular pattern in that family. It learns patterns from an input image and produces a larger image in the same style. It applies to both 2D tilemaps and 3D environments. It is not exactly AI — it is a constraint satisfaction problem — but in 2026 LLMs sit on top of it, making "request a dungeon with the mood you want in natural language" possible.

Houdini PCG turned SideFX's Houdini into the standard tool for game environment procedural generation. Ubisoft, Naughty Dog, Epic, and almost every large studio uses it. After 2025, with AI features added to Houdini, "an LLM node sitting on top of a node graph" has become a normal pattern.

Unreal PCG Framework is the procedural generation system Epic officially built into Unreal Engine starting in 5.4. It is node-based and offers a workflow similar to Houdini, but integrated directly into the engine. You can procedurally build grasslands, city streets, and dungeons inside the engine without external tools.

All three systems are absorbing AI nodes. Typing the mood of a grassland routes through an AI that auto-tunes PCG parameters. In spring 2026 this is early, but the direction is clear.

14. Game Engines plus AI — Unity, Unreal, Godot

Unity entered the AI category in earnest in 2023. Unity Muse bundled texture generation, sprite generation, and code generation. As of 2026 Muse is used more in AAA than among indies. Pricing is on a separate license.

Unity Sentis is a model inference runtime. You can run ONNX-format ML models directly inside the game. NPC behavior AI, procedural content generation, and adaptive difficulty are the typical use cases.

Unreal Engine 5 plus MetaHuman has become the standard in the character category. MetaHuman Creator builds photorealistic human characters, and layering NVIDIA ACE or Inworld's NPC AI on top has become a standard workflow. PCG Framework landed officially in Unreal 5.4, and 5.5 and 5.6 are bringing in more sophisticated AI integration.

Godot 4.4 is the underdog of the open-source camp, but it is closing the gap quickly. Community AI plugins are abundant, and GDExtension lets you embed Python ML libraries directly into a game. For indie developers, "free plus open" remains a powerful weapon.

A common pattern shows up in all three engines — a shift from "build outside in a SaaS and import" toward "generate inside the engine." Tool integration is going deeper.

15. Pixel Art AI — PixelLab.io and Scenario Pixel Mode

Pixel art is one of the hardest categories for general AI image generation. SDXL and Flux cannot pin the pixel grid precisely — anti-aliasing creeps in and pixels blur.

PixelLab.io is a pixel-art-only AI tool. It produces output that hits 8 by 8, 16 by 16, 32 by 32, and 64 by 64 grids exactly. It specializes in character sprites, tilemaps, and UI icons. Pricing offers a free tier plus paid subscriptions.

Scenario pixel mode was added in 2025. On top of Scenario's existing workflow integration, it adds pixel-art output, giving indie pixel game developers both character style learning and pixel output in a single tool.

The pixel art category is small in market size, but very active among indies. A large share of the Steam indie chart consists of pixel art games, and that category has long suffered from an artist shortage. AI tools are stepping into that gap.

16. Level Design AI — Promethean and Polycam

Level design — the layout and flow of a game's environment — has remained the area that needs the most human hand. That fact has not changed dramatically in 2026, but assistive tools have arrived.

Promethean AI is the most aggressive tool in environment design. It autoplaces furniture in a room, buildings on a street, and rooms in a dungeon. The human designer simply does fine adjustments on top — "this piece thirty centimeters to the left."

Polycam world capture is a different approach. It scans real space and brings it in as a game asset. An indie developer might scan a cafe nearby and layer NPCs on top. It is particularly strong in realistic games.

Apart from these two, Unreal PCG Framework, Houdini, and Unity's World Building Toolkit have become the standard tools for level designers. AI is the assist on top.

17. Photogrammetry and AI Texture Workflows

Another way to build 3D models is photogrammetry. Capture an object from many angles, then reconstruct a 3D mesh from those photos.

AI enters this workflow at two stages.

First, AI cleanup. Photogrammetry output is always noisy — missing patches, broken topology, rough surfaces. AI automatically cleans it up. NVIDIA Omniverse's AI tools are strong in this category.

Second, Polycam plus Trellis textures. Capture geometry with Polycam, then enhance textures with Trellis or another AI tool. This pattern has become common. Lighting noise at the time of capture is smoothed away by AI textures.

This workflow is especially strong in realistic games — that is, games that recreate real spaces. Cyberpunk-style and open-world genres adopted it quickly.

18. Cost at a Glance — From Indies to AAA

In spring 2026 the cost of game AI tools, at a glance:

Baseline for a one-person indie:

  • Scenario at twenty dollars a month (2D art)
  • Meshy at twenty or Tripo at twenty a month (3D models)
  • Inworld free tier (NPC AI prototyping)
  • ElevenLabs at five to twenty-two dollars a month (voice and sound effects)
  • Suno at ten dollars a month (music)

For about fifty-five to seventy-two dollars a month a single developer gets a full-stack AI workflow.

Mid-sized studio (around ten people):

  • Scenario Team at around two hundred a month
  • Meshy Ultimate at one hundred and twenty
  • Inworld usage-based (hundreds per month)
  • ElevenLabs Pro at ninety-nine
  • Substance 3D plus Firefly Materials at around fifty

About six hundred to one thousand a month covers the core tooling. Set against ten salaries, that is essentially negligible.

AAA studio:

  • An in-house LLM fine-tune (buy or rent a cluster of NVIDIA H100 or H200 GPUs)
  • Inworld Enterprise, or build your own NPC AI
  • Substance Enterprise
  • Houdini plus custom AI nodes
  • Omniverse licenses

This scales to hundreds of thousands or millions per year. But against saved outsourcing costs and time-to-market value, the ROI is clear.

19. Legal and Ethical Issues — Steam AI Labels, Voice Actor Unions, Licensing

In September 2024, Steam updated its policy on AI-generated content. Game registrations must disclose both "pre-generated AI content" (assets made with a trained AI model) and "live-generated AI content" (content the AI produces while the game runs). AI labels appear on the game page. Buyers see those labels before deciding.

The policy stirred a heated debate among indie developers. Some pushed back with "there is nothing to be embarrassed about using AI, why slap a label on us." Others said "we welcome the transparency." The data is showing that the labels themselves do not materially affect sales. Players still buy games that are fun, label or not.

Voice actor union concerns are a separate issue. SAG-AFTRA made AI voice usage a central issue in the 2024 to 2025 video game voice actor strike. The settled contract included a clause requiring explicit consent from a voice actor before their voice can be used in AI form. Replica Studios follows this model most cleanly — every voice model in its library sits on top of a licensed agreement with the voice actor.

Asset licensing is the issue of training data. Game-specific tools like Scenario and Layer AI are relatively safe because they fine-tune on data that the user uploads. But base models such as SDXL and Flux remain opaque about their training data, and that is still a legal risk. A few AAA studios sidestep that risk by adopting only "models trained on clean-license data," like Adobe Firefly or NVIDIA Edify.

20. Korean Game Studios plus AI

Between 2024 and 2026, the Korean game industry invested heavily in AI.

NCsoft unveiled its in-house LLM, VARCO, in 2023. It is a Korean-specialized LLM used for NPC dialogue, translation, and QA inside NCsoft games. It is also licensed to external partners. From 2025 onward, VARCO has expanded into multimodal, covering image and voice as well.

Krafton, the operator of PUBG, runs a game AI lab investing in LLMs, NPC AI, and generative content. Pilot projects integrating AI NPCs into in-house games are underway.

Smilegate invests in AI through its Future Lab. It has already piloted AI-based content generation in several mobile titles.

Pearl Abyss, the operator of Black Desert, develops AI tools internally alongside its proprietary engine. There are reports that AI workflows are being aggressively introduced into its new IP development.

Nexon and Netmarble are on similar trajectories. Nexon has expanded AI hiring alongside its US expansion through a subsidiary. Netmarble runs in-house LLM research.

The Korean game industry's strength is mobile live operations, and AI is going deep into that domain — ops automation, user support, automatic content generation. The Korean pattern is "put AI into live operations" more than "put AI into AAA console games."

21. Japanese Game Studios and AI — Square Enix, Capcom, Nintendo

The Japanese game industry approaches AI differently from Korea.

Square Enix's president officially stated in his 2024 New Year message that the company would aggressively use AI. Joint research with NVIDIA is ongoing, and the company invests in NPC AI and content generation for its own games.

Capcom builds AI tools internally alongside the RE Engine. Public announcements have confirmed AI in some text and translation workflows. That said, the company has stated that core IP — Monster Hunter, Resident Evil — keeps character assets human-built.

Bandai Namco owns the most diverse IP portfolio, and uses AI in proportionally many ways. AI shows up in content generation for several casual titles.

Sega invests in AI alongside its subsidiary ATLUS. Automation has entered RPG content generation, translation, and QA.

Nintendo is the most conservative. It has not publicly denied using AI, but it has not announced any AI use in core IP — Mario, Zelda, Pokemon. The Nintendo design philosophy itself is built on "the perfect handcrafted human finish."

miHoYo is a Chinese company, but Japan is a major market. Genshin Impact, Honkai, and Zenless Zone Zero are reported to integrate AI tools into their workflow. Character assets, however, remain artist-led.

22. The Future — Infinite Open Worlds, Real-Time Texture Generation, Fully AI Games

To close, here is the direction we can read from spring 2026 looking ahead two or three years.

1. AI-generated levels spreading. If Promethean, Houdini, and Unreal PCG are today's assistive stage in environment design, the next stage is "AI generates entire levels." An indie developer types "a stranded space base in a barren desert" once and gets a playable level in an hour.

2. Infinite open worlds. If No Man's Sky used procedural generation to create an infinite universe, the next step is AI populating that universe with "characters, cities, and cultures that all differ." A single player never sees the same city twice in a lifetime. AAA studios have started pouring resources into this category.

3. Real-time texture generation. Inside the game, when a player makes their own asset — a self-crafted sword, a self-styled outfit — AI generates the texture in real time. This category is growing fast on UGC platforms such as Roblox and Fortnite.

4. Fully AI games. The most distant future, but "one person generates a whole game from a single prompt" is becoming less science fiction. An indie designer types "a 1990s Japanese RPG, thirty hours long, mage protagonist" and a playable game emerges in a week. As of spring 2026 we are still a way off, but the industry consensus is that this lands somewhere between 2028 and 2030.

5. Settling ethical and legal norms. The three issues — AI labels, voice actor licensing, training data provenance — are settling into clearer norms. The EU, US, Korea, and Japan are each crafting different rules, and international game releases ride the intersection of those rules.

23. If You Have to Pick One Tool

Closing this article, here is a recommendation.

For a one-person indie: Start with Scenario (2D), Meshy (3D), the free tier of Inworld (NPC), ElevenLabs (voice), and Suno (music). Sixty to seventy dollars a month buys a full-stack AI workflow. A prototype lands within a month.

For a mid-sized studio of five to twenty people: Use Layer AI or Scenario Team for the 2D art pipeline, Meshy Ultimate or Tripo for 3D, Inworld Pro or Convai for NPCs, and Substance 3D plus Firefly Materials for textures. A monthly tooling spend of around one thousand dollars replaces outsourcing of more than ten thousand.

For a AAA studio: Keep external SaaS for prototyping only, and run core workflows on an in-house LLM fine-tune plus your own NPC AI. Ubisoft NEO NPC and the Square Enix and NVIDIA collaboration are the proof points for that direction. Evaluate NVIDIA Omniverse plus Edify 3D plus ACE as a full stack.

In every category: read the situation as "AI tools multiply human work speed by ten" rather than "AI tools replace humans." Great art directors, writers, and game designers are still needed. But the volume of output those people can produce per hour, with the right tools in hand, is now categorically different.

Epilogue — A Second Spring for Game Development

In spring 2026, game development is enjoying its second spring. The first one was the explosion of PC and console gaming from the late 1990s into the early 2000s. This second one — right now — is the era in which a single person on a laptop can produce AAA-grade content.

In this spring, indie developers ship games at scales that were not possible before. AAA studios ship games two or three times larger with the same headcount. The biggest beneficiary is the mid-sized studio in between — close to AAA in content volume, close to indie in cost.

The names of the tools will shift over time. Scenario could be replaced. Inworld could be acquired by a larger company. Microsoft Trellis's next version could become the new standard. But the underlying current does not reverse. AI has entered every step of game development, and the ceiling on what one person can ship has moved, permanently.

In 2026 we make our next games on top of that new ceiling. Where that ceiling goes next is up to us.

References

Discuss on Twitter • View on GitHub