Latent Rewards for Video Generation Models скачать в хорошем качестве

Latent Rewards for Video Generation Models

In this AI Research Roundup episode, Alex discusses the paper: 'Video Generation Models Are Good Latent Reward Models' The paper tackles how to better align video generation models with human preferences using reward feedback learning directly in the latent space. Instead of relying on costly vision–language models that operate on fully denoised RGB frames, the authors propose a Process-Aware Video Reward Model (PAVRM) that evaluates noisy latent representations across timesteps. PAVRM reuses early blocks of a pretrained video generator to extract spatiotemporal features from noisy latents, compressing them into a fixed token for timestep-aware quality assessment. Building on this, Process Reward Feedback Learning (PRFL) uses PAVRM as a latent-space reward during training, enabling more efficient, temporally dense supervision that can correct early motion and structural errors. Paper URL: https://arxiv.org/abs/2511.21541 #AI #MachineLearning #DeepLearning #VideoGeneration #RewardModeling #PreferenceLearning #DiffusionModels #GenerativeModels