GAN-Engine

Runtime	Docs / License	Tests

GAN-Engine is a general-purpose adversarial research environment. It began with super-resolution and is expanding into inpainting, conditional and unconditional image generation, and text-guided synthesis. The architecture is intentionally modular: if your data can be described as channels (plus optional masks or prompts), GAN-Engine lets you normalise it, condition on it, and model it without rewriting the training loop.

What makes it different?

• Task agility. Switch from super-resolution to inpainting or prompt-driven generation by choosing a different YAML preset.
• Composable architectures. Swap generator or discriminator families, layer in attention, register new modules, or wire up diffusion-ready adapters without touching trainer code.
• Rich conditioning. Mix class labels, segmentation masks, temporal context, or text embeddings to guide generation.
• Perceptual losses that understand your bands. Assign which channels feed into VGG, LPIPS, CLIP, or custom extractors so hyperspectral, medical, or grayscale data still benefit from perceptual guidance.
• Robust training loop. Warm-up phases, adversarial ramps, mixed precision, gradient accumulation, EMA, and Lightning 1.x/2.x compatibility are baked in.
• Experiment telemetry. Automated logging, validation panels, and checkpoint summaries keep researchers and stakeholders in sync.

These docs track the roadmap toward multi-task generative modelling. Expect super-resolution instructions alongside previews of inpainting, conditional, and text-to-image tooling. Each release adds templates and references for the new capabilities.

Choose your adventure

Goal	Where to start
Install the toolkit	Getting started
Understand the module layout	Architecture
Configure experiments	Configuration
Prepare datasets	Data
Launch training	Training + Training guideline
Run inference or sampling	Inference
Inspect Lightning internals	Trainer details
Explore benchmarks & samples	Results

Typical workflow

1. Duplicate a template config. Everything starts with YAML: pick a task, define the dataset loader, normalisation, conditioning signals, and architecture.
2. Point to your data. Use the built-in dataset selectors or register your own loader for DICOM, NIfTI, Zarr, tiling pipelines, COCO annotations, or prompt datasets.
3. Train with confidence. Kick off python -m gan_engine.train --config <config.yaml> and monitor metrics in Weights & Biases or TensorBoard.
4. Validate and export. Swap in EMA weights, export to ONNX/TorchScript, tile huge scenes, or sample from unconditional models without leaving the CLI.

Ecosystem

GAN-Engine is designed as the backbone of a broader generative stack. Companion repositories will host task-specific assets (prompt datasets, mask generators, diffusion adapters) while this core focuses on reusable infrastructure. Contributions that expand the ecosystem are welcome!