Free Stable Diffusion on Google Colab

Run AUTOMATIC1111's Stable Diffusion WebUI for free on Google Colab — no local GPU required.

---

Features

• Realistic Vision V6 — photorealistic model out of the box
• One-click setup — 3 cells, ~6 minutes, zero configuration
• Pre-configured settings — DPM++ SDE + Karras, 30 steps, CFG 5, Hires fix (768×512 → 1200×800)
• ADetailer — auto-fixes distorted hands & faces after every generation
• Cloudflare Tunnel — reliable public URL (replaces Gradio's unstable --share)
• Multi-fallback cloning — handles Stability-AI's private repo with automatic fork discovery
• Python 3.12 compatible — patches tokenizers build issues automatically
• API enabled — call from Python, curl, or any HTTP client

Quick Start

1. Click the Open in Colab badge above, or go to Google Colab → File → Upload notebook → select stable_diffusion_colab.ipynb
2. Runtime → Change runtime type → Runtime type: Python 3, Hardware accelerator: T4 GPU → Save
3. Runtime → Run all — sit back for ~6 minutes while it installs, downloads the model, and launches
4. Click the trycloudflare.com URL that appears in the Cell 3 output → paste a prompt → Generate!

Pre-configured Defaults

Setting	Value
Model	Realistic Vision V6
CLIP Skip	2 (critical for this model)
Sampler	DPM++ SDE
Schedule	Karras
Steps	30
CFG Scale	5
Base resolution	768 × 512
Hires fix	Enabled → 1200 × 800
Upscaler	R-ESRGAN 4x+
Denoising	0.40
Batch count	4
ADetailer	Enabled (hands + faces)
Face restoration	CodeFormer
Negative prompt	Pre-filled with anti-artifact terms

Settings are enforced through three config files (config.json, ui-config.json, params.txt) before launch — the WebUI opens ready to go.

Fixing Hands & Faces

Stable Diffusion can still occasionally produce distorted hands. This notebook addresses it multiple ways:

1. ADetailer extension — pre-installed, auto-enabled. Detects hands/faces after generation and repaints them at higher detail using hand_yolov8n.pt and face_yolov8n.pt.

2. Negative prompt — always include:

   bad hands, extra fingers, fused fingers, too many fingers, missing fingers, malformed hands, bad anatomy
   

3. Positive prompt (for scenes showing hands):

   detailed hands, correct finger anatomy, five fingers
   

API Usage

The --api flag is enabled by default. Once the WebUI is running:

import requests, base64
from PIL import Image
from io import BytesIO

url = "https://your-url.trycloudflare.com/sdapi/v1/txt2img"

response = requests.post(url, json={
    "prompt": "mountain landscape at sunset, cinematic, 8k",
    "negative_prompt": "blurry, low quality, bad anatomy",
    "steps": 30,
    "cfg_scale": 5,
    "override_settings": {"CLIP_stop_at_last_layers": 2},
    "width": 768,
    "height": 512,
}).json()

img = Image.open(BytesIO(base64.b64decode(response["images"][0])))
img.save("output.png")

Full interactive API docs: https://your-url.trycloudflare.com/docs

How It Works

The private repo problem

The Stability-AI/stablediffusion GitHub repo (required by AUTOMATIC1111) went private. This notebook solves it with a 3-tier fallback strategy:

1. Known public forks — tries Hafiidz and camenduru forks first
2. GitHub API — queries fork list sorted by stars
3. GitHub Search — broader search as a last resort

After cloning, the notebook validates the repo structure (checks for ldm.modules.midas) and saves the working URL + commit hash. Cell 3 sets environment variables so launch.py skips all git operations (no fetch from private URLs).

Other fixes baked in

Issue	Fix
tokenizers build failure on Python 3.12	Relaxes transformers==4.30.2 pin to >=4.30.2
Gradio tunnel hangs	Uses Cloudflare Tunnel (cloudflared) instead of --share
Missing taming module	Auto-clones CompVis/taming-transformers and symlinks
PyTorch version conflicts	Uninstalls default Colab torch, installs matched cu121 set

Troubleshooting

Problem	Fix
All clone URLs fail	Find a public fork at Stability-AI/stablediffusion/forks and add its URL to Cell 1
Out of memory	Reduce to 512×512, disable Hires fix, or lower batch count
WebUI won't load	Ensure T4 GPU is enabled. Factory reset runtime → re-run all cells
Distorted hands	Check ADetailer is enabled. Add hand terms to negative prompt
Slow first image	Normal — first generation compiles CUDA kernels (~30-60 sec)

Full reset: Runtime → Factory reset runtime → Change runtime type → T4 GPU → Run all cells

Limitations

• Free Colab sessions last ~12 hours, then everything resets — save your images
• The trycloudflare.com URL changes every session
• T4 GPU has 15 GB VRAM — batch size >1 with Hires fix may OOM
• SD 1.5-based models cannot render legible text in images

License

This notebook is provided as-is for educational purposes. It uses:

• AUTOMATIC1111/stable-diffusion-webui (AGPL-3.0)
• Bing-su/adetailer (AGPL-3.0)
• Realistic Vision V6 (CreativeML Open RAIL-M)
• Stable Diffusion v1.5 (CreativeML Open RAIL-M) — base architecture

---

If this saved you time, give it a ⭐!