A client brings 15 photos of their product — a new sneaker model. They need to place them on ad mockups: beach, mountains, studio. Basic Stable Diffusion doesn't know this object — the result depends on a random seed. If the seed isn't fixed, each prompt will yield a different angle, color, texture. DreamBooth solves the problem: it fine-tunes the model on 5–20 images, memorizing a unique identifier for the subject (e.g., sks sneaker). We use this approach for brand avatars, characters, and artistic styles. Experience with SDXL, LoRA, ControlNet ensures generation quality without overfitting. By combining DreamBooth, LoRA, and prior preservation, we achieve high-quality personalization without overfitting. The team has 5+ years of experience in CV and NLP, having completed over 100 fine-tuning projects. Typical cost ranges from $500 to $2000 depending on complexity.
DreamBooth is a method proposed by Google Research for fine-tuning text-to-image models to a specific subject.
Subject Preservation with DreamBooth
DreamBooth ties a rare token sks to the visual features of the object via prior preservation loss. This prevents "language drift" — the model does not forget general class concepts (e.g., 'sneakers' in general). Prior preservation loss uses class images (e.g., 'sneaker' without the subject) to keep the model from forgetting what ordinary sneakers look like. This is implemented by random sampling from the pre-trained model. Result: the subject is recognizable in any context.
Technically, the process consists of two stages: dataset preparation and training LoRA weights. LoRA (Low-Rank Adaptation) freezes the original SD weights and adds adapters — this requires 3x less VRAM than full fine-tuning (8 GB vs 24+ GB). LoRA is up to 10x faster and uses 3x less VRAM than full fine-tuning.
Why LoRA Is More Efficient Than Full Fine-Tuning
| Parameter | LoRA DreamBooth | Full Fine-Tuning |
|---|---|---|
| VRAM (SDXL) | 8–12 GB | 24+ GB |
| Training time (500 steps) | 15–30 min | 2–4 hours |
| File size | ~150 MB | ~6 GB |
| Overfitting | Minimal | Common |
| Style combination | Yes (LoRA merging) | No |
LoRA is the production standard: fast deployment, small size, easy to combine with other LoRAs (e.g., style + subject). Using LoRA reduces compute costs by up to 70%, saving approximately $500 per project on average. Each generation uses 30 inference steps and guidance scale 7.5 for optimal quality.
How to Prepare a Dataset for DreamBooth
The first thing an engineer encounters is the quality of the source images. The model copies angles, lighting, background. If all shots are taken in the same studio — DreamBooth will learn the studio as part of the subject.
- Image collection. Need 10–20 shots from different angles (front, side, top), different lighting (natural, artificial). The subject should occupy 50–80% of the frame. Avoid heavy occlusion (hand, shadow).
- Cropping and centering. Bring all images to a square of 1024x1024. Use the function from the listing below.
- Augmentation. To improve generalization, apply random horizontal flip, slight rotation (up to 10°), brightness/contrast changes. Strong distortions break geometry.
- Segmentation (optional). If the subject is a person, use RMBG 2.0 for isolation.
- Prior preservation. Generate 100–200 class images (e.g., 'sneaker' without the subject) using the base model. These images are used in the prior preservation loss.
from PIL import Image
import os
def prepare_dreambooth_dataset(
source_images: list[str],
output_dir: str,
target_size: int = 1024
) -> None:
os.makedirs(output_dir, exist_ok=True)
for i, img_path in enumerate(source_images):
img = Image.open(img_path).convert("RGB")
# Center and crop to square
width, height = img.size
min_dim = min(width, height)
left = (width - min_dim) // 2
top = (height - min_dim) // 2
img_cropped = img.crop((left, top, left + min_dim, top + min_dim))
img_resized = img_cropped.resize((target_size, target_size), Image.LANCZOS)
img_resized.save(f"{output_dir}/{i:03d}.jpg", quality=95)
print(f"Prepared {len(source_images)} images in {output_dir}")
Training: Choosing Hyperparameters
The Diffusers script for SDXL is run via accelerate. We recommend --mixed_precision="fp16" and --use_8bit_adam to save memory. The LoRA rank (r=64) balances adaptation and generalization.
accelerate launch train_dreambooth_lora_sdxl.py \
--pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0" \
--instance_data_dir="./training_images" \
--output_dir="./dreambooth_output" \
--instance_prompt="a photo of sks person" \
--resolution=1024 \
--train_batch_size=1 \
--gradient_accumulation_steps=4 \
--learning_rate=1e-4 \
--lr_scheduler="constant" \
--lr_warmup_steps=0 \
--max_train_steps=500 \
--seed=42 \
--mixed_precision="fp16"
More details about the script can be found in the official Diffusers documentation.
Key hyperparameters:
| Parameter | Range | Comment |
|---|---|---|
| Steps | 200–1000 | >1000 — risk of overfitting |
| Learning rate | 1e-4 to 1e-5 | Lower = more stable, but slower |
| Batch size | 1–2 | Limited by VRAM |
| Prior preservation | Yes | Use 100–200 class images |
The optimal number of steps depends on the subject's complexity. For simple objects (product on white background) 300-500 steps is enough. For complex ones (person with clothing details) — up to 800-1000. It's better to start learning rate at 1e-4 and decrease with cosine schedule.
If after training the model generates only one angle or ignores the background — it's a sign of overfitting. Solution: increase prior preservation weight, reduce steps, add augmentation.
Integration and Production Deployment
After training we get LoRA weights (usually ~150 MB). Load them into a custom StableDiffusionXLPipeline:
from diffusers import DiffusionPipeline
import torch
def train_dreambooth_sdxl(
instance_images_dir: str,
instance_prompt: str,
class_prompt: str,
output_dir: str,
num_steps: int = 800,
learning_rate: float = 1e-4
) -> str:
import subprocess
result = subprocess.run([
"accelerate", "launch", "train_dreambooth_lora_sdxl.py",
"--pretrained_model_name_or_path", "stabilityai/stable-diffusion-xl-base-1.0",
"--instance_data_dir", instance_images_dir,
"--instance_prompt", instance_prompt,
"--class_prompt", class_prompt,
"--output_dir", output_dir,
"--max_train_steps", str(num_steps),
"--learning_rate", str(learning_rate),
"--resolution", "1024",
"--train_batch_size", "1",
"--gradient_checkpointing",
"--mixed_precision", "fp16",
"--use_8bit_adam",
], capture_output=True)
return output_dir
def generate_with_dreambooth(
lora_path: str,
prompt_template: str,
subject_token: str = "sks"
) -> bytes:
pipe = DiffusionPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
torch_dtype=torch.float16
).to("cuda")
pipe.load_lora_weights(lora_path)
prompt = prompt_template.replace("{subject}", subject_token)
image = pipe(prompt, num_inference_steps=30, guidance_scale=7.5).images[0]
import io
buf = io.BytesIO()
image.save(buf, format="PNG")
return buf.getvalue()
After training, LoRA can be combined with ControlNet for precise control over pose, depth, or edges. For example, set a character's pose via OpenPose while keeping the appearance trained with DreamBooth.
Our Work Process
When ordering fine-tuning, we perform the following steps:
- Task analysis — we study your references, define the subject's class, choose the base model (SD 2.1, SDXL, or SD 3).
- Dataset preparation — we help with cleaning and augmenting images.
- LoRA training — we select hyperparameters, run training, check for overfitting.
- Testing — we generate 50+ variants in different contexts, pick the best checkpoint.
- Deployment — we deploy the model to cloud infrastructure (SageMaker, Vertex AI) or deliver files for local use.
- Documentation and support — we provide API docs, inference examples, and one month of support.
Timeline: from 2 days for simple objects to 3 weeks for a character with animation (sequential LoRA). Cost is calculated individually after assessment.
Common Mistakes and How to Avoid Them
- Overfitting — model generates only one angle. Solution: reduce steps, increase prior preservation, add augmentation.
- Wrong token — using a common word (e.g.,
person) leads to mixing with other subjects. Choose a rare token likesks. - Small dataset — fewer than 5 images won't let the model learn the object. Minimum 10.
- Poor background — if background is not diverse, the model ties the subject to one environment. Use shots with different backgrounds.
What's Included
- Prepared and augmented dataset (up to 20 images)
- Trained LoRA model (~150 MB file)
- Checkpoint with best quality (selected from 50+ generations)
- API documentation and inference example in Python
- Cloud deployment (SageMaker/Vertex AI) on request
- One month of technical support
Contact us for a consultation on your project. Get an estimate of timelines and cost — write to us, and we'll prepare a proposal within a day.
Order model fine-tuning to get a consistently recognizable subject in any context.
Stable Diffusion DreamBooth fine-tuning with LoRA delivers high-quality personalization. For best results, use 10-20 diverse images and a rare token like sks. This approach ensures the subject is preserved across contexts without overfitting.