Exploring the frontier of AI data, evaluation, benchmarks, and open-source infrastructure.
LEVI: Can AI Image Generators Preserve the Details That Define Visual Identity?
LEVI evaluates visual identity fidelity in AI image generation across 1,375 tasks in six domains. The best evaluated configuration achieves a 58.69% pass rate.
A New Benchmark for Spatial Reasoning: Can VLMs Imagine How Space Changes After an Object Moves?
MindEdit-Bench tests whether VLMs can predict spatial changes after objects move or rotate, using 1,003 verified questions from 120 indoor scenes.
Beyond Trial and Error: SolveEdit Benchmarks Visual Problem Solving in Generative AI
SolveEdit evaluates visual problem solving across 2,728 cases and shows how explicit planning improves AI image editing accuracy and content preservation.
AlignVid: Taming Visual Dominance via Training-Free Attention Modulation in Text-guided Image-to-Video Generation
ZODA-contributed work AlignVid was accepted to ICML 2026. It introduces a training-free, inference-time attention modulation method that enables image-to-video models to follow text-based editing instructions more accurately.
2026Show, Don't Tell: Morphing Latent Reasoning into Image Generation
ZODA-contributed work LatentMorph was accepted to ICML 2026. It integrates implicit latent-space reasoning into text-to-image generation, allowing models to dynamically refine their outputs during the generation process.
2026ScalingAR: Scaling Confidence for Autoregressive Image Generation
ZODA-contributed work ScalingAR was accepted to ICML 2026. It proposes a test-time scaling framework for autoregressive image generation that leverages the model’s own confidence signals to improve generation quality and efficiency.
2026