Mix Forcing Pre-Training Enhances Interactive Video World Models

Researchers have introduced Mix Forcing, a novel causal pre-training method designed to improve interactive video world models. This technique, implemented with Mixture-of-Bidirectional-and-Autoregressive (MoBA) attention, trains models using both teacher forcing and bidirectional training to preserve visual priors while enabling autoregressive generation. The approach also incorporates a chunk-wise cross-attention mechanism for online text interaction, conditioning video chunks on environment and local prompts without revealing future instructions. Experiments show that models pre-trained with Mix Forcing achieve state-of-the-art visual quality and temporal accuracy, even over extended generation periods. AI

IMPACT Enhances the quality and temporal accuracy of interactive video generation models.

RANK_REASON The item is an academic paper detailing a new method for training AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• alphaXiv
• arXiv
• CatalyzeX
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• Influence Flower
• Mix Forcing
• Mixture-of-Bidirectional-and-Autoregressive (MoBA)
• SANA-WM-Bench
• ScienceCast
• WBench

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →