Bird Image Generation - Master's Thesis

Comparative study of deep generative models for fine-grained bird image synthesis using the CUB-200-2011 dataset (11,788 images, 200 species).

---

Models

#	Model	Type	Resolution
1	DCGAN	Unconditional GAN	64×64
2	cDCGAN	Class-conditional GAN	64×64
3	Stable Diffusion v1.5 + LoRA	Text-guided diffusion	256×256

---

Dataset

CUB-200-2011 - Caltech-UCSD Birds-200-2011
11,788 images across 200 fine-grained bird species.

Methods

1. DCGAN - Unconditional

Standard Deep Convolutional GAN without class conditioning. Uses soft label smoothing and evaluates FID / IS every 10 epochs over 150 epochs of training.

Architecture:

• Generator: 5-layer transposed convolution stack, input noise dim = 100, output 64×64×3
• Discriminator: 5-layer convolution stack with LeakyReLU

2. Conditional DCGAN (cDCGAN)

Class-conditional extension. The Generator concatenates a 100-dim noise vector with a learned 200-dim class embedding. The Discriminator projects the class label as an extra spatial channel (1×64×64) concatenated to the image input.

3. Stable Diffusion v1.5 + LoRA

Fine-tunes Stable Diffusion v1.5 on CUB-200-2011 using LoRA adapters (rank 4) applied to UNet cross-attention layers (to_q, to_k, to_v, to_out). Training uses HuggingFace PEFT + Accelerate with FP16 mixed precision over 20 epochs.

---

Evaluation

Metrics are computed using torchmetrics:

Metric	Description
FID	Fréchet Inception Distance - measures distributional similarity between real and generated images (lower is better)
IS	Inception Score - measures quality and diversity of generated images (higher is better)

For statistical analysis, each model is evaluated over 20 independent runs (500–1000 images per run). Results are saved to CSV for downstream hypothesis testing.

---

Requirements

pip install torch torchvision torchmetrics
pip install diffusers transformers peft accelerate
pip install pytorch-fid torch-fidelity torchviz

---

Usage

Open main.ipynb in Google Colab (GPU recommended). The notebook is organized into the following sections:

1. Setup - Mount Google Drive, extract dataset, install dependencies
2. DCGAN - Unconditional model training
3. cDCGAN - Conditional model training
4. Stable Diffusion + LoRA - Fine-tuning, inference, and epoch-by-epoch evaluation
5. Statistical Evaluation - 20-run FID & IS measurement for all models

---

Results

Trained model checkpoints and metric logs are saved during training. CSV output files:

File	Contents
metrics_log.csv	Per-epoch G/D loss, FID, IS, duration (GAN models)
lora-sd15-cub200/training_metrics_SD.csv	Per-epoch loss, FID, IS (Stable Diffusion)
fid_and_is_scores_dcgan.csv	20-run FID & IS for DCGAN
fid_and_is_scores_cdcgan.csv	20-run FID & IS for cDCGAN
fid_and_is_scores_stablediffusion.csv	20-run FID & IS for SD + LoRA

---

Would you like to see my master's thesis?

Feel free to contact me. The thesis is written in Polish.