Paincontrol: identity-preserving pain expression transfer with generative diffusion models

Automated pain detection models rely on large, diverse facial expression datasets; yet pain expression data are scarce due to privacy concerns and ethical constraints. This study explores synthetic pain expression generation as a potential solution, investigating whether expression transfer methods are sufficiently accurate to expand training sets by transferring expressions from a limited set of real data to a virtually unlimited set of identities. This work represents an initial step toward the use of privacy-preserving generative AI for synthetic data generation to improve in pain detection. We introduce PainControl, a novel facial landmark-guided method for identity-preserving pain expression transfer. Rather than using motion signals from video, as in prior methods, our method adapts a ControlNet approach, using text and dense facial landmarks to control fine-grained facial muscle activations while maintaining identity consistency through image embeddings. Identity preservation enables the generation of demographically diverse datasets, improving generalizability across populations. We compare PainControl to existing expression transfer methods and evaluate (i) perceptual realism, (ii) identity preservation, (iii) facial action unit (AU) transfer accuracy, and (iv) downstream pain detection performance. Our model produces realistic synthetic pain expressions, outperforming baselines in human-rated Likert-scale assessments. However, AU transfer analysis reveals challenges in accurately synthesizing expression intensities-particularly for AU4 (Brow Lowering) and AU43 (Eye Closure), which are crucial for pain recognition. When used to augment real pain datasets, synthetic images did not improve classifier performance, likely due to artifacts, AU misalignment, and the lack of temporal motion cues. Critically, however, our experiments in data-scarce regimes, where real pain expressions are extremely limited, demonstrate that synthetic augmentation through PainControl provides significant and consistent performance gains for pain detection models. These results validate that our ControlNet-based approach has reached sufficient maturity and accuracy to serve as a reliable data augmentation tool in pain detection domains, particularly in scenarios, where acquiring real data are prohibitively difficult or ethically constrained. This finding establishes the practical viability of expression transfer methods for addressing the fundamental data scarcity challenge in affective computing and clinical pain assessment. These results highlight both the promise and current limitations of generative models for pain expression, pointing to future research in AU alignment and temporal modeling for clinical-grade applications.