How it works

ComfyUI exposes a REST API. The bot sends your prompt
to ComfyUI via websocket, waits for the image, then
sends it back to Telegram. That's the whole loop.

async def generate_image(prompt: str) -> bytes:
    # Build ComfyUI workflow with your prompt
    workflow = build_workflow(prompt)

    # Send to ComfyUI API
    async with websockets.connect(COMFYUI_WS) as ws:
        await queue_prompt(workflow, ws)
        image_data = await wait_for_output(ws)

    return image_data

async def handle_message(update, context):
    prompt = update.message.text
    await update.message.reply_text("Generating... ")

    image = await generate_image(prompt)
    await update.message.reply_photo(image)

The tricky parts

1. ComfyUI workflow as JSON
ComfyUI workflows are JSON graphs. Export any workflow
from the UI, save as a template, then inject the user
prompt at runtime. This lets you swap models without
changing bot code.

2. Queue management
Multiple users sending prompts simultaneously = queue
needed. I use asyncio queues to process one generation
at a time, with the user getting a "position X in queue"
message.

3. GPU access without paying
For development: Kaggle notebooks (30hrs GPU/week free),
Google Colab free tier.
For production hosting the bot itself (not the GPU),
you need almost nothing — it's just a Python process
making API calls.

Good free/cheap options:

Railway.app — free tier, deploy straight from GitHub, zero config

Render.com — free tier, auto-deploys on git push

Fly.io — free allowance, great for bots

Any $4/month VPS — DigitalOcean, Hetzner, Vultr

The bot uses ~80MB RAM idle. A $4 Hetzner ARM server
handles it perfectly.

Full source code

Everything is open source:
👉 Github

rahulkhunte / ai-image-bot

🤖 Jenerator Bot - AI Image Generation

Production-ready Telegram bot with GPU-accelerated image generation using Stable Diffusion XL Turbo. Generate high-quality images in under 4 seconds with natural language prompts.