Definition
Plain language
A training trick that teaches an AI to keep going even when the recent past it's building on is a bit wrong.
As stated in the literature
Training regimen that applies independent, per-frame random noise levels to context tokens so a diffusion model learns to predict from partially corrupted history, improving long-horizon rollout stability.
Why it matters: Without it, small mistakes pile up over a long sequence until the output falls apart, so it keeps generation stable over time.
For example, an AI generating a long video learns to keep producing sensible frames even when the earlier frames it's relying on came out slightly blurry or off.
Heard on the show
“The method is called diffusion forcing, and the image is a musician.”Episode 206 — How Four-Second Clips Become Hours of Playable AI Soccer