Report
ID-Forcing is claimed to enable minute-scale video generation without fine-tuning
HuggingPapers says the method aligns KV caching and conditioning with training to address drift beyond autoregressive video models’ training horizon.
TLDR
HuggingPapers describes ID-Forcing as a way to keep long autoregressive video generation in-distribution. It says the models drift beyond their training horizon and claims that aligning KV caching and conditioning with training enables minute-scale videos without fine-tuning.
Combined views
4.4K
2 Sources, first seen ago
