Show HN: Getting AI Models to Wink – The Wink Test (www.cinemodels.ai)

The Wink Test is a lightweight, hands-on benchmark and demo that probes whether image-to-video generative models can perform a very specific, localized instruction: make a subject wink with the correct eye (or any eye). The web interface runs multiple backends (examples shown: veo3, Kling, Gen-3, Seeddance, Moonvalley, Wan 2.2, Minimax Hailuo-02, Midjourney), lets users view and show prompts, randomize seeds, play/sync outputs, request new tests, and subscribe to benchmark alerts. Results are ranked so you can compare models’ success at following this fine-grained, spatially precise command. For the AI/ML community this is useful because the task exposes real weaknesses and differences in multimodal models’ controllability, compositional instruction-following, and semantic understanding of left/right and local facial attributes. As a minimal, reproducible test it highlights issues in alignment, prompt engineering, stochasticity from seeds, and model-specific biases—while also offering an easy way to track progress across releases. Practically, it’s a concise probe for improving training data, conditioning mechanisms, and evaluation metrics for video generation, and it surfaces downstream concerns such as reliability for identity-preserving edits and potential misuse in deepfake scenarios.