Ian's MIT-licensed Ultra Elements packs 39 real-time spell effects for the browser, from fireballs to portals, and its plan shows how parallel AI agents built each element under a 3 ms frame budget.
WhiteTowerAI's Apache-2.0 OhMyGame lets an agent design, generate assets for, code, playtest and publish a game. Beta 5 adds bring-your-own coding agents, an MCP server, Godot, Unity and Unreal plugins, and interactive stories.
Korrocorp released Kinloom 1.0 under MIT: one-file clients for Unity, Godot, Unreal via C++, Rust, Lua and more that add talking NPCs, per-frame decisions and generated content, validate every model answer against your example, and fall back to it offline.
Roblox's Q3 2026 shareholder letter says over 80% of its top 100 games now use AI tools, Build has passed 500,000 users and 70,000 published games, and NPC playtesters, 3D asset diffusion and the Roblox Reality world model are next.
SEELE AI put the EVA01 Unified All-Task generation adapter on Hugging Face: a native 3D model that turns a prompt or one image into a GLB through a Qwen3-VL-conditioned SS, Shape-512, Material-512 cascade built on TRELLIS.2.
Patch's MIT-licensed pixelfx ships 232 seeded, recolorable pixel art VFX written as code, sprite-sheet export for Godot, Unity and Phaser, and a Claude Code skill that drafts an effect, renders it, looks at the frames and refines it.
Scenario's MIT-licensed 3D OS gives Claude Code, Codex or Cursor six departments and about 80 skills, from image-to-3D, rigging and retopology to game-ready export, plus expert families that drive Blender, Maya, ZBrush, Unreal, Unity, Godot and Roblox Studio on your machine.
Moonshot AI's new 3D Make gives Kimi skills and plugins to script parametric models in Blender and CAD, call Tripo for organic meshes, and assemble scenes and small games in Unreal, Unity or Godot, live now on Kimi.ai and Kimi Work.
Ramen's Aura has released Autopilot under Apache-2.0: an editor plugin for Unreal 5.4–5.8, Unity 6 and Godot 4.7+ that takes a one-line prompt, plays the game in real time with real inputs, learns per project and writes a bug report. Aura's own UE 5.8 test has it passing 7 of 7 tasks to Claude Code's 5.
Blocksfield, an MIT-licensed entry in Comfy's #OpenSourceHiggsfield challenge, has a local Qwen3.8-27B build a real Roblox place in Studio, play it, judge the screenshots and repair it, while ComfyUI makes the art, 3D props, music and trailer on the same GPU.
camenduru's new MIT-licensed OpenCartridge puts sprite-sheet turnarounds, Loop Hero pixel art, CorridorKey mattes and Idol image-to-3D on one local canvas, each running as a frozen Docker worker on your own NVIDIA GPU.
New open ComfyUI-Idol custom nodes and a LichtFeld Studio plugin wrap Sperid Labs' Idol, turning one image into a native 3D Gaussian splat in about 30 seconds on a 7.2 GB VRAM peak.
Tsinghua's HairDeck generates stylized anime-style hair cards plus the spring-bone chains that animate them from a single image, then exports GLB or a full VRM avatar. Code, weights and a free demo are live.
YC Labs and Olam Labs ran swarms of Claude and GPT agents in Minecraft, Bed Wars and Mindustry, then open-sourced the harness and environments. Here's what the agents did and how to run your own.
Speridlabs' Idol turns a single render or photo into a 3D Gaussian splat of up to a million Gaussians. Weights and code are Apache 2.0, it peaks at 7.2 GB VRAM, and there's a free Hugging Face Space to try it today.
Tripo's new Agentic Canvas turns a chat request into a reusable node workflow that generates, assembles, rigs and exports 3D assets, from a reference image to an FBX character or a UE5 scene project. It's in beta for Max and Team users first.
Meshy's 3D generation is now built into the open-source MCP for Blender plugin, so Claude Code, Codex or Cursor can turn prompts and concept art into textured models that land in Blender at real-world scale, for 30 credits each.
NVIDIA's MotionBricks now has its own open repo with a pretrained SOMA human model, ONNX exports and a playable scene demo. It generates real-time character motion from target poses, and the weights allow commercial use.
Google DeepMind's Genie 4 generates explorable worlds in real time with native audio, characters that speak, and live text steering through Gemini. It opens as a trusted-tester preview with a Project Genie waitlist, and Gemini API access is coming.
Scenario's new Pixel VFX Forge and Claude skill turn a description like "frost nova" into a layered pixel-art effect, rendered locally at no cost, with sprite sheets, a JSON atlas, GIFs and a Godot 4 pack.
WanGP's new One World Generator adds a Worlds gallery where you build a world from text with Tencent's Hy-Nomas 8B or from an image with EVOKE Turbo 14B, walk it with the keyboard, save the run, and record it as video, starting at 8GB of VRAM.
A new open-source node pack, ComfyUI-Hy-Nomas, runs Tencent's Hy-Nomas world model natively in ComfyUI on 12GB NVIDIA cards, adds start-from-image worlds and a WASD drive mode, and ships five ready workflows. Tencent also posted an official BF16 Diffusers build.
NVIDIA shipped RTX Remix 1.6 on October 23. The modding Toolkit swaps its old AI Tools for ComfyUI 2.0 texture and mesh workflows, adds a headless host and modding-agent skill with a new Streamable HTTP MCP endpoint, and brings DLSS 3D-Guided Neural Rendering controls, up to 6x frame generation, Windows on ARM, and an open-source Minecraft Java 1.21.1 beta mod.
NVIDIA released Nemotron 3.5 Bolt E4B 4GB, an open Gemma-4-based small model tuned for in-game NPCs and interactive characters that mix dialogue with tool calls. It runs in about 3.5GB of VRAM with 548 ms time to first token on an RTX 5070, ships as a GGUF under OpenMDW-1.1, and NVIDIA demoed it as a blacksmith trader that crafts and hands over items on its own.