Dreambooth-Stable-Diffusion

Fine-tunes a pre-trained Stable Diffusion model using 3-5 user-provided images of a specific subject by learning a unique token embedding while preserving general image generation capabilities through class-prior regularization. The training process uses PyTorch Lightning to optimize the text encoder and UNet components, employing a dual-loss approach that balances subject-specific learning against semantic drift via regularization images from the same class (e.g., 'dog' images when personalizing a specific dog). This prevents overfitting and mode collapse that would degrade the model's ability to generate diverse variations.

Automatically generates synthetic regularization images during training by sampling from the base Stable Diffusion model using class descriptors (e.g., 'a photo of a dog') to prevent overfitting to the small subject dataset. The system iteratively generates diverse class-prior images in parallel with subject training, using the same diffusion sampling pipeline as inference but with fixed random seeds for reproducibility. This creates a dynamic regularization set that keeps the model's general capabilities intact while learning subject-specific features.

Saves and restores training state (model weights, optimizer state, learning rate scheduler state, epoch/step counters) to enable resuming interrupted training without loss of progress. The implementation uses PyTorch Lightning's checkpoint callbacks to automatically save the best model based on validation metrics, and supports loading checkpoints to resume training from a specific epoch. Checkpoints include full training state, enabling deterministic resumption with identical loss curves.

Provides a configuration system for managing training hyperparameters (learning rate, batch size, num_epochs, regularization weight, etc.) and integrates with experiment tracking tools (TensorBoard, Weights & Biases) to log metrics, hyperparameters, and artifacts. The implementation uses YAML or Python config files to specify hyperparameters, enabling reproducible experiments and easy hyperparameter sweeps. Metrics (loss, validation accuracy) are logged at each step and visualized in real-time dashboards.

Selectively updates only the text encoder (CLIP) and UNet components of Stable Diffusion during training while freezing the VAE decoder, using PyTorch's parameter freezing and gradient masking to reduce memory footprint and training time. The implementation computes gradients only for unfrozen parameters, enabling efficient backpropagation through the diffusion process without storing activations for frozen layers. This architectural choice reduces VRAM requirements by ~40% compared to full model fine-tuning while maintaining sufficient expressiveness for subject personalization.

Generates images at inference time by composing user prompts with a learned unique token identifier (e.g., '[V]') that maps to the subject's learned embedding in the text encoder's latent space. The inference pipeline encodes the full prompt through CLIP, retrieves the learned subject embedding for the unique token, and passes the combined text conditioning to the UNet for iterative denoising. This enables compositional generation where the subject can be placed in novel contexts described by the prompt (e.g., 'a photo of [V] dog on the moon') without retraining.

Orchestrates the training loop using PyTorch Lightning's Trainer abstraction, handling distributed training across multiple GPUs, mixed-precision training (FP16), gradient accumulation, and checkpoint management. The framework abstracts away boilerplate distributed training code, automatically handling device placement, gradient synchronization, and loss scaling. This enables seamless scaling from single-GPU training on consumer hardware to multi-GPU setups on research clusters without code changes.

Implements classifier-free guidance during inference by computing both conditioned (text-guided) and unconditional (null-prompt) denoising predictions, then interpolating between them using a guidance scale parameter to control the strength of text conditioning. The implementation computes both predictions in a single forward pass (via batch concatenation) for efficiency, then applies the guidance formula: `predicted_noise = unconditional_noise + guidance_scale * (conditional_noise - unconditional_noise)`. This enables fine-grained control over how strongly the model adheres to the prompt without requiring a separate classifier.

Loads pre-trained Stable Diffusion model weights (v1.4, v1.5, or compatible checkpoints) and initializes the text encoder, UNet, and VAE components with proper architecture matching and weight initialization. The implementation validates checkpoint compatibility by verifying layer names and dimensions, handles different checkpoint formats (safetensors, PyTorch pickle), and supports loading from local paths or Hugging Face model hub. This abstraction enables seamless model swapping without modifying training or inference code.

Preprocesses input subject images by resizing to 512x512 (or specified resolution), applying center cropping or padding to maintain aspect ratio, and normalizing pixel values to [-1, 1] range for VAE encoding. The pipeline includes optional augmentation (random crops, flips) during training to improve generalization, and deterministic preprocessing during inference. Images are encoded to VAE latent space (4x downsampled, 64-dimensional) before diffusion training, reducing memory footprint and enabling efficient batch processing.

Computes a weighted combination of two loss terms during training: (1) subject loss on personalized images (MSE between predicted and actual noise in diffusion process) and (2) regularization loss on class-prior images (MSE on synthetic class images). The total loss is: `loss = subject_loss + lambda * regularization_loss`, where lambda is a hyperparameter controlling the regularization strength. This dual-loss formulation prevents overfitting by penalizing the model for degrading its ability to generate diverse class examples while learning subject-specific features.

Executes the image generation process through iterative denoising steps, starting from random noise and progressively refining the image by predicting and subtracting noise at each timestep. The pipeline applies text conditioning (via CLIP embeddings) and classifier-free guidance at each step, using a scheduler (e.g., DDPM, PNDM) to determine noise levels and step sizes. The implementation batches conditioned and unconditional predictions for efficiency, applies guidance interpolation, and decodes the final latent representation through the VAE to produce the output image.