Reference Forcing

Long Video Generation with Region-Guided Drift Correction

Two-hour videos

Each video is generated in one pass, without restarts or stitching. Shown here at 60× speed; the clock gives the time in the generated video.

0:00:00
Otter · An adorable happy otter stands on a surfboard wearing a yellow lifejacket, riding along turquoise tropical waters near lush tropical islands.
0:00:00
Tokyo walk · A stylish woman in a black leather jacket and a long red dress walks down a bustling Tokyo street filled with warm glowing neon.
0:00:00
Robot · A robot with sleek silver armor and glowing blue circuits explores a neon-lit cyberpunk city.

Comparison with long-video methods

A 60-second video for the same prompt and the same first chunk. All methods are built on LongLive and attend to 12 latent frames.

Prompt · A woman walks briskly through a bustling, wet city during heavy rain, carrying an umbrella and wearing a dark coat. Reference Forcing keeps her face and the strap on her shoulder as in the first frames, while the baselines lighten the face and widen the strap.