VintageGAN: Controllable Vintage Image Synthesis

VintageGAN is an academic portfolio project for learning controllable vintage image degradation. The user-facing experience is intentionally similar to a photo editor: choose a preset or tune six sliders for grain, scratches, dust, vignette, color shift, and blur. Under the hood, the project trains a conditional image-to-image model to imitate procedural vintage targets.

This repository is a research prototype. It does not claim production quality, published metrics, or trained checkpoints until those artifacts are generated from reproducible runs.

Current Goal

Run locally first on an RTX 3050 4GB class GPU at 256x256.
Keep the same code path extendable to 384x384 and 512x512 cloud GPU runs.
Report only metrics produced by saved experiment artifacts.
Keep presets and manual slider values explicit and reproducible.

What Is Implemented

Conditional U-Net style generator with a dynamic bottleneck condition projection.
Conditional PatchGAN discriminator with a dynamic patch grid.
Procedural vintage target generator for six independent defect dimensions.
Explicit preset vectors in configs/presets.yaml.
Local/cloud training profiles in configs/training_config.yaml.
Mixed precision and gradient accumulation hooks for memory-constrained GPUs.
Inference CLI/API that accepts presets or manual slider values and can write metadata JSON.
Dataset loading with explicit train/val split folders to avoid leakage.

What Is Not Yet Claimed

No real FID/SSIM/PSNR numbers are claimed in this README.
No “production ready” status is claimed.
No trained checkpoint is bundled.
ImageNet/FilmSet use is not assumed unless access and license are verified.

Installation

Use Python 3.10 or 3.11. For the local RTX 3050 profile:

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements-local-cu118.txt
pip install -e .

For CUDA 12.1 cloud images:

pip install -r requirements-cloud-cu121.txt
pip install -e .

Dataset Layout

Use a legally allowed clean-image dataset and split it explicitly:

data/public_images/
  train/
    image_000001.jpg
  val/
    image_000001.jpg

Before reporting results, record the dataset source and license in configs/training_config.yaml under dataset.dataset_name and dataset.dataset_license.

Training Profiles

The default profile is local_256, intended for a 4GB RTX 3050:

python training/pretrain.py --config configs/training_config.yaml --profile local_256
python training/gan_train.py ^
  --config configs/training_config.yaml ^
  --profile local_256 ^
  --generator-checkpoint checkpoints/generator_pretrain_best.pth

Cloud profiles use the same scripts:

python training/pretrain.py --config configs/training_config.yaml --profile cloud_384
python training/pretrain.py --config configs/training_config.yaml --profile cloud_512

Each training run writes a metadata manifest under outputs/experiments/.

Inference

Preset mode:

python inference/apply_filter.py input.jpg outputs/input_vintage.jpg ^
  --checkpoint checkpoints/generator_final.pth ^
  --preset warm_film ^
  --image-size 256 ^
  --metadata-output outputs/input_vintage.json

Manual slider mode:

python inference/apply_filter.py input.jpg outputs/custom.jpg ^
  --checkpoint checkpoints/generator_final.pth ^
  --grain 0.6 --scratch 0.2 --dust 0.4 --vignette 0.5 --color-shift 0.7 --blur 0.1

Python API:

from inference import VintageFilter

vintage = VintageFilter(checkpoint="checkpoints/generator_final.pth", image_size=256)
output = vintage.apply("input.jpg", conditions="dusty_archive")
output.save("outputs/dusty_archive.jpg")

Presets

Preset vectors live in configs/presets.yaml.

The canonical condition order is:

grain, scratch, dust, vignette, color_shift, blur

Current academic/demo presets:

soft_fade
warm_film
dusty_archive
scratched_negative
heavy_vintage

Backward-compatible aliases such as light, medium, and heavy are also kept for older scripts.

Evaluation Plan

Metrics should be generated only after training:

SSIM/PSNR against synthetic procedural targets, with clear caveats.
FID only when a real, licensed vintage reference set is available.
Condition-control accuracy only when defect detectors are trained and validated.
Ablations for no adversarial loss, no perceptual loss, and no consistency loss.

Recommended portfolio evidence:

Config file.
Git commit.
Dataset source/license.
Hardware notes.
Training duration.
Saved sample grids.
Metrics JSON produced by evaluation scripts.

Tests

pytest tests -v

Minimum checks expected before publishing results:

import models, import training, and import inference pass.
Generator and discriminator shape tests pass at 256x256.
Dataset determinism and train/val split checks pass.
One tiny smoke training run completes.
A checkpoint can be loaded for inference.

Repository Structure

configs/      training profiles and preset vectors
defects/      procedural target generation
models/       generator, discriminator, attention, detector modules
training/     dataloaders, losses, pretraining, GAN training
evaluation/   metrics and ablation scaffolding
inference/    single-image and batch filter application
tests/        unit and integration tests

Citation

This project is not a publication. If you use it in coursework or a portfolio, cite the repository and include the exact commit hash used for results.

License

MIT. Dataset licenses are separate and must be checked for the images you train or evaluate on.

Downloads last month: -; Downloads are not tracked for this model. How to track