StyleGAN2 Eye Manipulation on CPU

Extended implementation of StyleGAN2 for eye feature manipulation on CPU
Based on StyleGAN2: Analyzing and Improving the Image Quality of StyleGAN

---

Key Features

• CPU-compatible StyleGAN2 implementation
• Eye feature manipulation (size, openness, distance, color)
• Pre-computed eye direction vectors
• Robust tensor dimension handling
• Direct image processing alternative

---

Quick Start

# Run the simplified notebook
.\run_eye_manipulation_simple.bat

# OR run the direct image processing script (fallback)
.\run_simple.bat

---

User Guide

For a complete step-by-step guide, see the USER_GUIDE.md. This guide provides detailed instructions for both beginners and advanced users.

---

Comprehensive Documentation

1. README_CPU_EYE_MANIPULATION.md: Complete project overview
2. TECHNICAL_DETAILS.md: Technical implementation details
3. PROCESS_DOCUMENTATION.md: Step-by-step process documentation
4. TROUBLESHOOTING.md: Solutions for common issues

---

Requirements

Tested on:

• PyTorch 1.3.1
• CUDA 10.1/10.2

---

Usage

1. Prepare LMDB Datasets

python prepare_data.py --out LMDB_PATH --n_worker N_WORKER --size SIZE1,SIZE2,SIZE3,... DATASET_PATH

This converts images to JPEG and pre-resizes them. You can create multi-resolution datasets with comma-separated --size arguments.

---

2. Training (Distributed)

python -m torch.distributed.launch --nproc_per_node=N_GPU --master_port=PORT train.py --batch BATCH_SIZE LMDB_PATH

• Supports Weights & Biases logging via --wandb.

---

3. SWAGAN Support

This implementation supports SWAGAN:

python -m torch.distributed.launch --nproc_per_node=N_GPU --master_port=PORT train.py --arch swagan --batch BATCH_SIZE LMDB_PATH

---

4. Convert Weights from Official Checkpoints

Clone the official StyleGAN2 repo and download .pkl checkpoints.

Convert weights:

python convert_weight.py --repo ~/stylegan2 stylegan2-ffhq-config-f.pkl

---

5. Generate Samples

python generate.py --sample N_FACES --pics N_PICS --ckpt PATH_CHECKPOINT

Set --size if you trained at a different resolution.

---

6. Image Projection

python projector.py --ckpt [CHECKPOINT] --size [GENERATOR_OUTPUT_SIZE] FILE1 FILE2 ...

---

7. Closed-Form Factorization

See Closed-Form Factorization:

Extract eigenvectors:

python closed_form_factorization.py [CHECKPOINT]

Apply extracted directions:

python apply_factor.py -i [INDEX_OF_EIGENVECTOR] -d [DEGREE_OF_MOVE] -n [NUMBER_OF_SAMPLES] --ckpt [CHECKPOINT] [FACTOR_FILE]

Example:

python apply_factor.py -i 19 -d 5 -n 10 --ckpt [CHECKPOINT] factor.pt

Will generate 10 random samples, with latents moved along the 19th eigenvector with degree ±5.

---

Implementation Process

This CPU-compatible eye manipulation implementation involved:

1. CPU Operation Implementation

   • Replacements for CUDA-specific operations
   • Simplified upfirdn2d and fused_leaky_relu
   • CPU-friendly noise injection

2. Dimension Handling Fixes

   • Fixed tensor dimension issues in latent vectors
   • Robust shape checking and validation
   • Automatic shape correction

3. Pre-Computed Eye Directions

   • Direction vectors for eye size, openness, distance, color
   • Support for combined feature manipulation
   • Normalized vectors for consistency

4. Alternative Solutions

   • Direct image processing for eye manipulation
   • Automated model downloading
   • Robust error handling and diagnostics

---

Key Files and Components

• Core Files:

  • eye_manipulation_simple.ipynb: Main notebook for StyleGAN2-based eye manipulation
  • simple_eye_manipulation.py: Direct image processing alternative
  • eye_directions.py: Pre-computed eye feature direction vectors
  • download_models.py: Download official StyleGAN2 weights
  • op/cpu_ops.py: CPU-compatible operations for StyleGAN2

• Support Files:

  • TROUBLESHOOTING.md: Guide for fixing common issues
  • setup_stylegan2_ada.py: StyleGAN2-ADA integration setup utility
  • run_eye_manipulation_simple.bat: Run the simplified notebook (Windows)
  • run_simple.bat: Run the direct image processing approach (Windows)

---

Original StyleGAN2 Samples

Sample from FFHQ (trained on 3.52M images)

Sample from MetFaces with Non-leaking augmentations (150,000 iterations, trained on 4.8M images)

Samples from Converted Weights

Sample from FFHQ (1024px)

Sample from LSUN Church (256px)

---

License

• Model details and custom CUDA kernel codes from NVIDIA StyleGAN2
• LPIPS code from PerceptualSimilarity
• FID Inception V3 from pytorch-fid