Dungeon Ultimate

An offline AI dungeon master with real 3D dice, full D&D mechanics, uncensored on-device image generation and voice input — your adventures never leave your machine.

English · Русский

Overview

Dungeon Ultimate is a fully on-device AI roleplay engine — a tireless dungeon master that writes your story, runs real tabletop rules, rolls physics dice, illustrates the scene and reads it aloud. Everything happens on your own NVIDIA GPU: no cloud, no API keys, no accounts, no content filters, and nothing ever leaves your PC.

It is a heavily extended fork of open-dungeon, rebuilt around a local text model, an uncensored local FLUX image pipeline, a D&D-style game engine with real 3D physics dice, and on-device voice input. Run the launchers and play at http://localhost:3000.

Why it's different

Most "AI dungeon" apps are a thin wrapper around someone else's cloud LLM — your prompts get logged, the model is censored, and there are no actual game mechanics behind the prose. Dungeon Ultimate flips all of that:

• It runs on your hardware. The story model, the image model and the speech model all load on your GPU. Pull the network cable and it still works.
• There are real rules. A deterministic D&D 5e engine resolves checks, combat and damage with a server-side CSPRNG — the narrator declares the action, the engine decides the outcome, so the AI can't cheat.
• The dice are real. A genuine 3D physics die tumbles across the scene (three.js + cannon-es) and is forced to land on exactly the number the engine already rolled.
• It's uncensored. A local, unfiltered text model plus a local FLUX image pipeline — with an optional abliterated text encoder for the images — mean unrestricted, adult storytelling and art, entirely your call and entirely private.

Features

Real 3D physics dice

• A genuine d20 built on @3d-dice/dice-box-threejs (three.js + cannon-es) tumbles across the scene with real physics.
• Rolls are honest — the deterministic engine rolls first with a Node crypto CSPRNG, then the on-screen die is pinned (1d20@N) to land on that exact value. No fudging, no re-rolls.
• The settled die is colour-tinted by outcome (gold crit, red fumble) and logged to the adventure journal.

D&D game mode

• Character sheet — six D&D 5e ability scores (STR / DEX / CON / INT / WIS / CHA), AC, level, XP and conditions.
• d20 ability checks — the narrator declares a check (ability + DC); the engine rolls d20 + modifier, with natural 20 always a crit success and natural 1 always a fumble.
• HP & death — characters track current/max HP and flip to a dead state when they hit zero.
• Turn-based combat — the narrator can spawn enemies, resolve attack rolls against AC and apply damage; foes are tracked per-encounter.
• Adventure journal — every roll, hit, drop and death is appended to a player-facing log that doubles as the engine's audit trail.
• Loot drops — enemies and chests grant inventory items with slots, rarity tiers and stat modifiers; each drop can be illustrated and the portrait reused via image2image.
• Game mode is per-chat and on by default — keep it on for a full RPG session, or toggle it off for freeform narrative play.

Uncensored on-device image generation

• Scenes are illustrated locally by a quantized FLUX.2-klein-4B pipeline — no cloud, no key, no filter.
• The narrator calls an image tool mid-story and the picture renders right inside the passage.
• Runs out of the box on the ungated FLUX.2-klein SDNQ weights (no token, no filter); an optional abliterated text encoder (opt-in, gated) drops the last content guard for fully unrestricted 18+ art.
• Ken-Burns animation on rendered images plus one-click retry.

Voice input

• Speak your action instead of typing — a mic button captures audio and transcribes it on-device.
• Powered by NVIDIA Parakeet-TDT-0.6B-v3 ASR (via onnx-asr + ONNX Runtime GPU), running locally with no upload.

And the rest

• Live token streaming — the narrator's prose streams into the chat word by word.
• Voice narration (TTS) — turns can be read aloud by a local text-to-speech server.
• One model on the GPU at a time — the text LLM unloads while images render and reloads on the next turn, so each gets the whole GPU.
• Editable prompts & per-chat settings — narrator prompt, image prompt, world, style, characters, response length, voice.
• 7 play languages — the narrator, action chips, suggestions and TTS all follow your chosen language (Russian, English, Spanish, French, German, Chinese, Japanese), switchable in-app. The UI chrome is Russian; image prompts stay English for FLUX.
• Portable Windows launchers — install.bat / run.bat / stop.bat; models, runtimes and caches stay on a non-system drive.

Requirements

• OS: Windows 10/11 (install.bat + run.bat set up a fully portable, self-contained install)
• GPU: NVIDIA, 12+ GB VRAM (RTX 40xx/50xx fully supported; 20xx/30xx/Pascal selectable in the installer). The installer pins matching CUDA wheels (cu126 / cu128) per GPU.
• Node.js: bundled — install.bat downloads a portable Node 22 runtime into the project folder
• Python: bundled — install.bat creates two embedded Python 3.11 environments (text/TTS and image)
• Disk: ~30+ GB for the embedded runtimes plus model weights
• Model weights — all auto-downloaded, nothing to provide by hand. install.bat clones the (public) image backend — the Qwen3-TTS engine is bundled right in the repo at servers/tts_engine.py, nothing private to clone — and downloads a reference voice pack into servers/voices/; on first launch every model pulls itself from Hugging Face — the Gemma 4 12B GGUFs (text), the FLUX.2-klein SDNQ image weights, the Qwen3-TTS voice model, and the Parakeet ASR model. All of those are ungated. The fully-abliterated FLUX text encoder lives in a gated HF repo, so it stays opt-in: point IMAGE_SERVER_DEFAULT_BACKEND=flux-uncensored at it once you have repo access + an HF token. You can also drop your own .mp3 reference clips into servers/voices/.

The app is engineered to stay self-contained: temp files, Hugging Face caches, Torch caches and model stores are all redirected onto the project drive — nothing is written to C: or the registry.

Quick start

1. Clone

   git clone https://github.com/timoncool/dungeon-ultimate.git
   cd dungeon-ultimate

2. Install — run install.bat, pick your GPU, and let it download the portable Node + Python runtimes, install dependencies and build the web app.

   install.bat
   

3. Run — start the text, image, TTS and web servers together.

   run.bat
   

   Your browser opens at http://localhost:3000. Stop everything with stop.bat.

How to play

Start a new story — pick a genre, say who you are, and the narrator writes the opening scene.

Pure storytelling — the narrator's prose streams in word by word, book-style.

• Create a chat, set the world/style or pick a character, then type or speak an action — the narrator streams a story turn.
• Turn on game mode for a full D&D session: ability checks roll a real 3D die, combat resolves against AC, HP and loot are tracked, and every result lands in the adventure journal.
• Hit the mic button to dictate your action; it's transcribed locally by Parakeet.
• Toggle narration to have turns read aloud.
• Edit the narrator / image prompts in the side panels to retune tone and art direction.

Other projects by timoncool

Project	Description
ACE-Step Studio	AI music studio — songs, vocals, covers, videos
VideoSOS	AI video production in the browser
Foundation Music Lab	Music generation + timeline editor
Qwen3-TTS	Portable text-to-speech with voice cloning
SuperCaption Qwen3-VL	Portable image captioning
civitai-mcp-ultimate	Civitai API as an MCP server
ScreenSavy	Ambient screen generator

Authors

• Nerual Dreming — Telegram | neuro-cartel.com | ArtGeneration.me
• Нейро-Софт — Telegram | portable neural-network apps

Acknowledgements

Built on open-dungeon by @newideas99 — the original local AI roleplay app this fork extends. Huge thanks for the foundation.

The 3D dice are powered by @3d-dice/dice-box-threejs. Speech recognition uses NVIDIA's Parakeet-TDT-0.6B via onnx-asr. Imagery is produced with FLUX.2.

Support the author

I build open-source software and do AI research. Most of what I create is free and available to everyone. Your donations help me keep creating without worrying about where the next meal comes from =)

All donation methods | dalink.to/nerual_dreming | boosty.to/neuro_art

• BTC: 1E7dHL22RpyhJGVpcvKdbyZgksSYkYeEBC
• ETH (ERC20): 0xb5db65adf478983186d4897ba92fe2c25c594a0c
• USDT (TRC20): TQST9Lp2TjK6FiVkn4fwfGUee7NmkxEE7C

Star History

License

MIT — same as the upstream project. Do whatever you want; attribution appreciated.