ComfyCrawler by sorryaboutyourcats
ComfyCrawler (read-only edition) by sorryaboutyourcats
Dungeon Creation Wizard
Fill in the blank to build your custom 3D dungeon world.
This page only plays back dungeons someone already made. Making your own needs ComfyUI and roughly 50–95GB of models running on your own machine.
🏰 Get ComfyCrawler github.com/sorryaboutyourcats/ComfyCrawlerAttaching a picture makes creating a dungeon take longer.
Right now pictures are set to Use the picture itself: FLUX Kontext draws that line's hero, weapon, foes or walls straight from your picture, so they look like it. With a picture on every line, that can take nearly twice as long.
To speed it up, open ⚙️ Options and, under Generation Settings, set Attachments to Describe the picture in words. Qwen3-VL then writes down what your picture shows and the art is drawn from those words instead – only seconds more, but the art matches the description rather than looking exactly like your picture.
All of this is theirs: the idea, the look, what the foes do, and which of a hundred prompt wordings survived contact with a real render.
You fill in four blanks on the main menu – a style, a hero, a weapon, a foe – and a stack of models running on this machine builds the rest: the walls, ceiling and floor, the hero and their swing, three different foes, the story that scrolls past while you wait, the names, the footsteps, the music, and (if you switch it on) a short film of the final blow.
Then you walk it. Every dungeon you finish making is kept in dungeon_sessions/ and can be started again from 📜 History for nothing, because the expensive part already happened. That window reads either way: as rows, or as tiles – each dungeon drawn as its own hero standing in it.
A play-only copy of ComfyCrawler. Every dungeon in the list started as four blanks – a style, a hero, a weapon, a foe – and a stack of models running on sorryaboutyourcats’ own PC built the rest: the walls, ceiling and floor, the hero and their swing, three different foes, the story that scrolls past before you start, the names, the footsteps, the music, and on some of them a short film of the final blow.
All of that was made ahead of time, so here there is nothing to wait for: pick one and walk it. The pictures and sounds are fixed, but the maze is not – it is laid out fresh every time you press Start, at whichever difficulty is picked, so the same dungeon plays differently twice.
Want one with your own four blanks in it? 🛠️ Make your own runs, under the list, has both ways to get one.
🤖 AI agents can play too: llms.txt explains the game and how.
Written by sorryaboutyourcats together with Claude, in Claude Code on Opus 5. One Python server and one JavaScript file, no build step and no framework beyond Tailwind off a CDN.
sorryaboutyourcats decided what it should be and played every build; Claude wrote the code and measured what came back.
None of these run in this page. They ran once per dungeon, on one NVIDIA RTX 3090 through ComfyUI, and what they made is what you are playing.
A Python 3.10 standard-library HTTP server on 127.0.0.1:5555 serves the page and owns every run. Each model above is reached through ComfyUI on 127.0.0.1:8188: graphs are posted to its /prompt, finished images and audio are collected from /history, and the loading bar is fed by its WebSocket. SageAttention is switched on for the cutscene when the node is installed.
No cloud API, no account, no key, no MCP server. The only thing this page talks to is your own machine, and nothing you type into it leaves.
It is the same page and the same game code as the full version, with generating switched off. There is no server behind it: just static files, a list of the dungeons, and one saved bundle per dungeon holding its pictures, sounds, music, story and names. The maze, the foes’ moves and the fights all run in your browser.
No account and no tracking. Your options, stars and which bosses you have beaten are kept in this browser only, and go nowhere else.
Every file the pipeline loads, and whether this machine’s ComfyUI can see it. Click one to open the folder it belongs in. Anything missing can be downloaded from the setup panel on the main menu.
Every dungeon here started as four blanks – a style, a character, a weapon and an enemy – and a stack of models drew, wrote, named, scored and voiced the rest. Two ways to get one with your own four in it:
ComfyCrawler is free and open source. It needs ComfyUI and a GPU with plenty of VRAM (developed on an NVIDIA RTX 3090; other GPUs untested), plus about 50GB of models – 95GB with the sound, music and ending cutscene too – and the setup screen downloads whichever of those are missing for you. A dungeon then takes a few minutes to build, and is yours to keep and replay.
No graphics card, or no interest in installing anything? Message me your four and I'll run them here. Anything goes for the style – a place, a mood, a single word, a name in “quotes”. Write them out in one go, like this:
That dungeon isn't in this showcase – it may have been taken down. Here are the ones that are.
Erase this saved dungeon?
Its textures, sprites, sound effects, music and narration are deleted from disk. This cannot be undone – making it again means generating it again.
Stop filming this dungeon's ending movie?
Stop this download?
Close out the dungeon you are playing?
Erase the dungeon you are playing?
The run ends and its textures, sprites, sound effects, music and narration are deleted from disk. This cannot be undone – making it again means generating it again.
ComfyCrawler
Play a saved dungeon – it's already generated!
YOU DIED!
The dungeon keeps what it kills.
⬆️ LEVEL UP!
You reached Level 2. Choose upgrade!
DUNGEON ESCAPED!
You climbed the stairs out and conquered the dungeon!
A/D to turn
SPACE to use
TAP TO USE