Video Generation Models: A Survey of Post-Training and Alignment
Survey maps video-generation post-training into supervised, distillation, preference, and inference-time alignment methods.
This survey reviews post-training and alignment methods for video generation models, which often fail to follow intent, stay temporally coherent, or meet physical and safety constraints after pretraining. It distinguishes implicit from explicit alignment and groups methods into supervised fine-tuning, self-training and distillation, preference- and reward-based training, and inference-time control. It also covers datasets, benchmarks, and evaluation practice. Open challenges include scalable rewards, long-horizon consistency, stability versus expressiveness, and safety-aware generation.
- Survey of post-training and alignment for video generation models.
- Separates implicit alignment from explicit alignment.
- Covers SFT, self-training and distillation, preference and reward, and inference-time methods.
- Reviews datasets, benchmarks, and evaluation practices.
- Open issues include reward design, temporal consistency, and safety-aware generation.
From the original article
1 min read at huggingface.co
Video generation has rapidly progressed from short, low-quality clips to high-resolution, long-duration sequences with complex spatiotemporal dynamics. Despite strong generative priors learned through large-scale pretraining, pretrained video models often fail to reliably follow human intent, maintain temporal coherence, or satisfy physical and safety constraints.
By Chaoyu Li, Xiaoyi Gu, Yogesh Kulkarni · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.