Video Generation Models: A Survey of Post-Training and Alignment

Video Generation Models. A Survey of Post-Training and Alignment. By Chaoyu Li and twelve others, from Arizona State University, Twitch, and seven other institutions. Published on the thirtieth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Video generation has moved quickly. It has gone from short, low-quality clips to long, high-resolution sequences with complex dynamics in space and time. Large-scale pretraining gives these models strong generative priors. Even so, a pretrained video model often falls short in three ways. It fails to follow human intent reliably. It fails to stay coherent over time. Or it fails to satisfy physical and safety constraints.

Fixing that is harder for video than it is for images or text, and the survey names four reasons. Errors accumulate over time, so a small mistake early in a clip can grow. Motion and appearance are coupled. Several objectives trade off against one another. And supervision for temporal properties is limited. Those difficulties motivate systematic post-training: strategies that adapt a pretrained model to what people want without retraining it from scratch.

The authors present this as the first comprehensive review of post-training and alignment in video generation models. They treat post-training as a unifying framework, and draw one central distinction across it: implicit alignment versus explicit alignment, according to how the alignment signal is enforced. From that vantage, existing approaches fall into four broad categories. Supervised fine-tuning methods come first. Then self-training and distillation methods. Then preference-based and reward-based methods. And finally inference-time methods, which act when the model is being used.

The taxonomy is meant to give one coherent view of how alignment signals shape model behavior, across training and deployment alike, so that methods which look unrelated can be placed on the same map. Beyond the methods themselves, the survey reviews the datasets, benchmarks and evaluation practices in common use. It also sets out the open challenges. Reward design that scales. Temporal consistency over long horizons. The trade-off between stability and expressiveness. And generation that is aware of safety.

A single question runs under the whole map: where does the alignment signal enter, and how is it enforced once it does. Fine-tuning on curated data, learning from a model's own outputs or a teacher's, optimizing against preferences and rewards, or steering at inference: each is a different answer to that question, and the split between implicit and explicit alignment sorts those answers by how the signal is enforced. The stated aim is a structured conceptual foundation, with practical guidance alongside it, for building video generation models that are more controllable and more reliable.