Controllable image synthesis
Diffusion and flow-based generation where the user keeps a handle on the result — layout, identity, and attributes that stay put when everything else changes.
Diffusion modelsControllable generationImage editing
Research
Research
Everyday photos, video and aerial imagery look like different worlds. The modelling problems underneath them are close to identical.
Image restoration
Recovering a clean, faithful image from a degraded capture — noise, blur, low light, compression and missing regions — across photos, video, RAW sensor data and scientific imaging. The restored result has to be right, not just look right.
DenoisingVideo restorationInpainting & outpaintingRAW imaging
Synthetic data for downstream tasks
What synthesis is ultimately for: training detection, segmentation and anomaly models when real labels run out — including applied settings such as industrial inspection and remote sensing.
Data augmentationAnomaly detectionSegmentationTransfer