Causal Generative AI
Harness the power of causal generative AI for super-realistic image synthesis—transforming how you develop, test, and validate AI.
Our technology enables high-fidelity data augmentation to boost performance, stress testing to uncover hidden vulnerabilities, and bias detection to ensure fair and equitable predictions across diverse populations.
For faster, safer, and better innovation in imaging AI.
Enhance your training datasets with clinically valid, diverse, and accurate synthetic scans. Our causal generative AI creates high-fidelity medical images that preserve causal relationships—boosting model performance and robustness.
Simulate edge cases, imaging artefacts, and underrepresented patient profiles to rigorously test your AI model's reliability. Our platform helps uncover hidden failure modes before deployment, supporting safer and more trustworthy clinical AI.
Expose and quantify performance disparities across demographic subgroups using controlled, diverse synthetic data. Our technology enables fairness audits and bias mitigation strategies, aligning your AI with regulatory and ethical standards.
Our approach
Our technology is based on rigorous science, theoretical analyses, and extensive experimentation, resulting in the most advanced methodology for targeted super-realistic image synthesis, combining causal modelling with the power of deep generative AI.
With a click of a button, we simulate ‘what-if’ scenarios: What would this patient's scan look like if there were no disease? What would the image characteristics be if a different scanner were used?
This is what we call Imaginable Imaging.
Peer-reviewed
Our work is peer-reviewed and published in the leading AI conferences and journals such as NeurIPS, ICML, ICLR, and Nature Machine Intelligence. Our technology is rooted in rigorous science, theoretical analyses, and extensive experimentation.
Check out some of our selected papers.
MICCAI, 2025
CF-Seg is a framework that enhances anatomical segmentation in medical images by generating pseudo-healthy counterfactuals. Using deep structural causal models, it improves segmentation accuracy without modifying existing models. Evaluations on chest X-ray datasets show CF-Seg outperforms standard methods, especially in cases with pleural effusion, and is preferred by radiologists.
ICML, 2025
This paper introduces diffusion-based causal mechanisms for generating high-fidelity image counterfactuals. By integrating semantic abduction and dynamic guidance into diffusion models, the framework balances identity preservation and causal faithfulness. It outperforms prior methods across tasks like face editing and medical artefact removal, advancing counterfactual reasoning in complex visual domains.
Medical Image Analysis, 2025
Counterfactual contrastive learning enhances robustness in medical image representation by generating realistic domain-shifted image pairs using causal generative models. Evaluated on chest X-rays and mammograms, it outperforms standard contrastive methods across diverse scanners and label regimes, improving generalization, especially for underrepresented domains and subgroups, using both SimCLR and DINO-v2 frameworks.
MICCAI, 2024
This paper identifies "attribute amplification" in counterfactual image generation, where unintended features change during interventions. We propose Soft Counterfactual Fine-Tuning (Soft-CFT), which uses soft labels to preserve non-intervened attributes. Experiments on chest X-rays show Soft-CFT improves faithfulness and reduces bias, enhancing reliability in medical imaging applications.
Nature Machine Intelligence, 2024
This paper presents a causal framework for understanding dataset bias in medical imaging. It identifies three bias types—prevalence, presentation, and annotation disparities—and shows how each requires distinct mitigation strategies. We propose a three-step fairness reasoning process and emphasize the need for explicit causal assumptions and transparency in model development and evaluation.
ICLR, 2023
This paper proposes a framework to evaluate counterfactual image models using axiomatic principles—composition, reversibility, and effectiveness—without needing ground-truth counterfactuals. It introduces metrics to assess model soundness, explores simulated interventions to mitigate confounding, and compares generative models across datasets, highlighting trade-offs in counterfactual fidelity and interpretability.
ICML, 2023
This paper introduces a deep causal generative framework for producing high-fidelity image counterfactuals. By combining hierarchical variational autoencoders with structural causal models, it enables estimation of direct, indirect, and total effects. The use of synthetic images could improve explainability, robustness, and fairness in imaging applications. The methodology is validated through effectiveness and composition metrics.
NeurIPS, 2020
This paper introduces Deep Structural Causal Models (DSCMs), a framework combining deep learning with structural causal models to enable tractable counterfactual inference. By leveraging normalizing flows and variational inference, DSCMs support all three levels of Pearl's causality ladder—association, intervention, and counterfactuals—demonstrated through synthetic and real-world imaging case studies.
Our team
We are a team of AI researchers from the Department of Computing at Imperial College London. Over the last five years, we have developed one of the most advanced methodologies for high-fidelity image synthesis powered by causal generative AI.
We have strong expertise in computer vision and medical imaging applications. And we love to collaborate.