Example results¶
This page showcases qualitative and quantitative examples from different domains to illustrate what GAN-Engine can deliver. Use these as inspiration when designing your own experiments.
Medical imaging¶
- MRI (T1-weighted, ×4): Generator
rrdb, perceptual loss LPIPS on channel 0, SAM weight 0.02. Produces crisp cortical boundaries and preserves contrast without introducing halo artefacts. - CT (lung, ×2): Generator
rcab, adversarial ramp over 15k steps, histogram matching enabled. Results maintain HU fidelity for radiologist review.
Remote sensing¶
- Sentinel-2 RGB-NIR (×4): Generator
lka, discriminatorpatchgan. Spectral angle mapper ensures vegetation indices remain stable. Suitable for agritech analytics. - SWIR minerals (×3): Custom generator with grouped convolutions and channel attention. Perceptual loss limited to bands 0–5 to respect domain-specific features.
Microscopy¶
- Fluorescence confocal (3D, ×2): 3D RRDB blocks with volumetric tiling. Model trained with total-variation regularisation to reduce ringing while enhancing fine structures.
- Histopathology WSI (×4): PatchGAN discriminator with feature matching. Results sharpen cellular boundaries and glandular textures for pathologist review.
Consumer photography¶
- Compressed JPEG (×4): ESRGAN baseline with stochastic residual blocks. Removes compression artefacts and restores detail in handheld shots.
- Drone imagery (×2): RCAB generator with Weights & Biases logging for on-site monitoring. Handles mixed lighting conditions.
Inpainting (roadmap)¶
- Urban facades: Mask-aware UNet generator guided by segmentation labels. Early experiments show seamless texture blending across occluded windows and signage.
- Satellite cloud removal: Sentinel-2 masks combined with conditional embeddings to reconstruct land-cover details beneath thin cloud layers.
Text-to-image & unconditional synthesis (roadmap)¶
- Creative posters: CLIP-conditioned generator fine-tuned on design references. Roadmap experiments target prompt alignment scores comparable to lightweight diffusion models.
- Material catalogues: Unconditional GAN trained on microscopy textures to sample novel grains for data augmentation.
Quantitative benchmarks¶
Super-Resolution¶
| Domain | Dataset | Scale | PSNR ↑ | SSIM ↑ | LPIPS ↓ | Notes |
|---|---|---|---|---|---|---|
| Medical | FastMRI knee | ×4 | 30.2 | 0.92 | 0.082 | LPIPS restricted to T1/T2 bands, EMA checkpoint |
| Remote sensing | SEN2NEON RGBNIR | ×4 | 28.7 | 0.89 | 0.115 | Histogram matching + SAM 0.05 |
Metrics are indicative and depend on configuration details, normalisation statistics, and training duration. Use them as starting points rather than hard baselines.
Inpaining¶
coming soon...
Unconditional Image Generation¶
coming soon...
text-to-image¶
coming soon...