Causal-Consistent Counterfactual Generation via Latent Interventions

Current disentangled representation approaches like StyleDiffusion and CookGALIP separate attributes but ignore causal relationships between them. This research introduces causal latent intervention, allowing users to generate counterfactual scenarios by modifying causes while maintaining realistic effects. For example, in food generation (building on CookGALIP), instead of just changing ingredients, you could ask "what would this dish look like if cooked at higher temperature?" and the model would realistically propagate the causal effects through texture, color, and composition. The key innovation is learning a causal graph over latent dimensions using techniques from causal discovery, then enabling do-calculus style interventions during generation. This differs from existing control methods by ensuring generated counterfactuals maintain causal consistency - addressing the semantic inconsistency problem that plagues current models. The approach could revolutionize applications from medical imaging (generating disease progression scenarios) to climate modeling (visualizing policy impacts).

References:

1. CookGALIP: Recipe Controllable Generative Adversarial CLIPs With Sequential Ingredient Prompts for Food Image Generation. Mengling Xu, Jie Wang, Ming Tao, Bing-Kun Bao, Changsheng Xu (2025). IEEE transactions on multimedia.
2. StyleDiffusion: Controllable Disentangled Style Transfer via Diffusion Models. Zhizhong Wang, Lei Zhao, Wei Xing (2023). IEEE International Conference on Computer Vision.