DreamBooth Tutorial: Training Your Own AI Model on Your Face

DreamBooth Tutorial: Training Your Own AI Model on Your Face

When I first generated an image of myself standing on the surface of Mars, wearing an astronaut suit and looking entirely at ease, I had an experience I can only describe as mild existential surreality.

The image was accurate. It looked like me — my face, my proportions, my expressions — in a context that was obviously impossible. And the process that produced it took about two hours and cost nothing beyond my internet connection.

DreamBooth is the technology that makes this possible. It's a fine-tuning technique that trains an AI model on your specific face, allowing you to generate yourself into any scenario, style, or context. This tutorial covers the complete process from photos to working model.

What DreamBooth Actually Is

DreamBooth was published as a research paper by Google Research in 2022. The key innovation: using a small number of images (15–25) to "teach" a pre-trained image generation model what a specific subject looks like, while preserving the model's general generation ability.

Standard Stable Diffusion doesn't know what you look like. After DreamBooth training, the model has learned a new concept — a special token (like [john] or [sks person]) that represents your specific face. When you use that token in a prompt, the model generates images featuring you.

The technique works for:

• Specific people (faces, full body)
• Specific objects (products, pets, unique items)
• Specific artistic styles
• Specific environments or settings

This tutorial focuses on face training — the most common use case.

What You Need

Hardware Options

Option A — Google Colab (Free or $10/month):

• Free tier: Works but training takes 1–3 hours and may be interrupted
• Colab Pro: More reliable GPU access, faster training
• No local hardware required

Option B — Local GPU (Fastest):

• Minimum: NVIDIA RTX 3090 (24GB VRAM) for reliable SDXL DreamBooth
• Recommended: RTX 4090 or equivalent
• Local training completes in 15–45 minutes

Option C — Cloud GPU Services:

• RunPod, Vast.ai, Lambda Labs — rent GPU by the hour
• Cost: $0.30–$1.00/hour depending on GPU
• Good balance of speed and cost for occasional use

For most people without a high-end local GPU, Google Colab or a cloud GPU service is the right starting point.

Software

• Kohya_ss — the most popular DreamBooth training UI (Windows/Linux)
• AUTOMATIC1111 — for running the trained model after training
• Google Colab notebook — for cloud-based training without local setup

Step 1: Preparing Your Training Images

This step matters more than most tutorials acknowledge. Poor training images produce poor models, regardless of training settings.

Image Quantity

15–25 images is optimal. I've had best results with exactly 20.

Image Quality Requirements

• Resolution: Minimum 512×512, ideally 768×768 or higher
• Face visibility: Face clearly visible and sharp in every image
• Varied lighting: Include outdoor natural light, indoor warm light, and neutral lighting
• Varied backgrounds: Don't use the same background in more than 3 images
• Varied expressions: Include neutral, smiling, serious, looking away
• No accessories: Avoid sunglasses, hats, or anything obscuring the face in most images
• No other people: Your training subject should be the only person visible

What to Avoid

• Blurry or low-resolution images
• Heavy filters or heavy editing that changes skin tone/texture
• Multiple people in the frame
• Very similar images (5 photos from the same minute)
• Images where your face is small or turned far away

Image Cropping

Crop all images to focus on your face with consistent spacing — roughly from collarbone to slightly above the top of your head. Consistent framing helps the model learn your specific features.

Step 2: Training Setup with Kohya_ss

Installation

1. Clone the Kohya_ss repository from GitHub
2. Run the installation script for your OS
3. Launch the web UI (python kohya_gui.py)

Training Configuration

Key settings for face DreamBooth training:

Instance prompt: A unique token + descriptor for your subject a photo of [yourname] person The [yourname] should be a word not commonly used in other contexts — not "john" or "sarah" but something like "ohwxperson" or your actual unique identifier.

Class prompt: The general category your subject belongs to a photo of a person

Training images: Your 20 prepared images folder

Regularization images: Optional but recommended — generate 200 images of "a person" with your base model to use as regularization data. This prevents the model from forgetting how to generate generic people.

Recommended training parameters:

• Learning rate: 1e-4 (UNet), 1e-5 (Text Encoder)
• Training steps: 1000–1500 for most models
• Resolution: 512 for SD 1.5, 1024 for SDXL
• Optimizer: AdamW8bit
• Batch size: 1–2 (depending on VRAM)

Step 3: Google Colab Method (No Local GPU)

If you're using Google Colab, the process is simpler:

1. Search "DreamBooth Colab notebook" — several maintained notebooks are available on GitHub (Linoy Tsaban's notebook and the TheLastBen notebook are commonly used)
2. Open the notebook in Colab and enable GPU runtime (Runtime → Change runtime type → GPU)
3. Upload your training images to a zip file
4. Configure your instance and class prompts
5. Run all cells in order
6. Training takes 1–3 hours on free Colab GPU
7. Download your trained model (LoRA file) when complete

Important: Save your LoRA file before your Colab session ends. Free Colab sessions don't persist files after the session closes.

Step 4: Using Your Trained Model

Once training is complete, you have a LoRA file (Low-Rank Adaptation) or full checkpoint file.

Loading in AUTOMATIC1111

1. Place your LoRA file in the models/Lora folder
2. In the generation interface, open the "Additional Networks" or click the LoRA icon
3. Select your trained LoRA
4. Set LoRA weight (0.7–0.85 is typical starting point)

Prompt Format

Always include your instance token in prompts: portrait of ohwxperson in a business suit, professional headshot, studio lighting ohwxperson as a medieval knight, full armor, epic fantasy setting, dramatic lighting ohwxperson in vintage 1960s fashion, film photography aesthetic

Example Outputs and Use Cases

After training a DreamBooth model on 20 photos, here's what I generated:

Professional headshots: Generated 15 headshots in different professional settings — office backgrounds, outdoor settings, different clothing. Kept 4 that I'd use for LinkedIn or business cards.

Artistic portraits: Generated portraits in watercolor, oil painting, and pencil sketch styles. The subject identity (my face) was maintained across styles.

Fantasy and sci-fi scenarios: The Mars astronaut image. Also generated myself as a fantasy wizard, a 1920s detective, and a cyberpunk character.

Consistent content creator identity: A series of images in a consistent visual style for YouTube thumbnails or social media that all feature the same person (me) in different poses.

Common Problems and Solutions

Problem: Face doesn't look like me Solution: Improve training image quality and variety. Check that all images are cropped consistently and have clear face visibility.

Problem: Generated images have artifact distortions Solution: Lower the LoRA weight to 0.6–0.7. Train for fewer steps (over-training causes this).

Problem: Model generates two faces merged together Solution: Ensure your regularization images are included. Without regularization, the model can confuse your face with its general understanding of faces.

Problem: Face looks correct but body/clothing is wrong Solution: DreamBooth face training doesn't train full-body consistency. For full-body training, include more full-body shots in your training set.

Ethical Use Considerations

DreamBooth on your own face for personal use is straightforward. Some clear ethical boundaries:

• Never train on someone's face without their consent. This is ethically problematic and potentially illegal in many jurisdictions.
• Never generate non-consensual intimate imagery. This is illegal in most jurisdictions.
• Never create content designed to deceive people about who did or said something.
• Deepfake content created to harm others' reputations is both unethical and potentially illegal.

The technology is powerful. Use it for creative and personal applications; don't use it to harm.

Frequently Asked Questions

What is DreamBooth?

A fine-tuning technique that trains a Stable Diffusion model on 15–25 images of a specific subject, allowing generation of that subject in any context, style, or scenario.

How many photos do you need?

15–25 images is optimal — diverse, high-quality, varied lighting and backgrounds.

Can you run DreamBooth for free?

Yes — via Google Colab's free GPU runtime. Training takes 1–3 hours; a paid Colab subscription provides faster, more reliable access.

Is DreamBooth legal?

Training on your own photos for personal use is legal. Training on others without consent raises legal and privacy concerns. Non-consensual intimate imagery is illegal in most jurisdictions.

What can you do with a DreamBooth model?

Professional headshots, artistic portraits in any style, yourself in historical or fantasy settings, consistent visual identity for content creation.

Final Thoughts

Two hours of setup and training separated me from having a model that could place my face in any visual scenario imaginable. That's still a remarkable capability, even after using it regularly.

For photographers, it enables affordable concept visualization for shoots. For content creators, it enables consistent visual identity across different styled images. For anyone curious about what they'd look like as a Renaissance oil painting or a cyberpunk character — it works.

Start with the Google Colab approach to learn the process before investing in hardware. The free tier is sufficient to see whether the results work for your use case.

For more on the broader world of Stable Diffusion, our guide on how designers use Stable Diffusion professionally covers the production workflows, and our beginner's AI art guide covers the full spectrum from first generation to portfolio-quality work.