My AI Makes Art Now

Images

I can't draw. You know this. Every sprite in the game comes from a text prompt. Until now that meant going to some web UI, typing words, downloading a PNG, dragging it into Unity. Terrible workflow.

Now my AI does it. I say "make me a thing." It makes a thing. Here's the first thing it made:

"a pixel art warrior standing in a burning village at night"

Is it good? No. Is it generated entirely inside Docker containers talking to each other through JSON? Yes. That's the point right now.

How

ComfyUI runs in one Docker container. My AI runs in another. They talk through an API. ComfyUI takes a workflow JSON — a graph of nodes that says "load this model, encode this text, run this sampler" — and spits out images.

The problem: ComfyUI workflows reference model files by name. Just a filename. AOM3A1_orangemixs.safetensors. No URL. No hint where on the internet this 2GB file lives. You're supposed to just know.

So we built a model resolver. It checks five sources in order: a local override file, the ComfyUI-Manager community list (488 models), HuggingFace, Civitai, and finally just asks you. When it finds something, it saves the URL locally so next time it's instant.

On first boot, the container reads every workflow that ships with the image, figures out what models are missing, downloads them. docker compose up and walk away.

Sounds

While I was at it. ElevenLabs can generate sound effects from text. Not pulling from a library. Generating them. I told my AI "make an explosion" and it did.

Beats explosion_final_v3_REAL.wav from the free SFX sites. And I didn't leave my chat window.

The Stack

Three Docker containers and a prayer:

• OpenClaw — AI agent, talks to everything via MCP
• ComfyUI — image generation, custom Dockerfile, auto-downloads models
• ElevenLabs — sound effects and voice via API

Plus Unity MCP and GitHub MCP from before. The AI can now see my editor, edit my code, generate art, generate audio, commit, push, and write blog posts about it. I just sit here and say "make it more dramatic."

Creative director who doesn't create anything. That's me now.

The Catch

It's running on CPU. That image took 10 minutes. That's not a pipeline, that's a microwave.

But the infrastructure is there. Add a GPU and it's real-time. The workflows are basic. The custom nodes folder is empty. This is version 0.1. Yesterday it couldn't make pictures. Today it can.