In-Distribution Forcing Enhances Long Video Generation at Test Time
| Source: HF Papers | Original article
Researchers introduce In-Distribution Forcing to enhance long‑video generation in autoregressive diffusion models, tackling drift problems such as color shifts, texture changes, and decaying motion dynamics.
A new test‑time technique called In‑Distribution Forcing (ID‑Forcing) promises to push autoregressive video diffusion models far beyond their usual short‑horizon limits. The method, unveiled in a pre‑print posted four days ago, works by reshaping the key‑value (KV) cache that underpins generation at inference time. Rather than retraining models, ID‑Forcing aligns KV caching and KV conditioning with the configurations used during training, using a “self‑caching” step that blocks out‑of‑distribution (OOD) entries from entering the rolling window. The result is a seamless extension of existing short‑horizon checkpoints – such as those from Self‑Forcing and LongLive – to minute‑scale video generation without any code or weight changes.
The advance matters because long‑form video synthesis has long been hampered by drifting, where colors, textures and motion dynamics gradually degrade. Prior work has relied on KV conditioning that only partially mitigates the issue. In the authors’ evaluations, ID‑Forcing not only held its own on standard video generation benchmarks but also delivered a marked reduction in drift, as confirmed by bespoke drift metrics and a user study. By eliminating the need for additional training, the approach could lower the barrier for developers to produce high‑quality, extended video content, opening doors for applications ranging from entertainment to simulation.
The community will now watch for broader adoption of the technique. Integration with other autoregressive video models, further large‑scale user studies, and real‑world deployments will test whether ID‑Forcing can become the default inference‑time recipe for long video generation. As we reported on Oct 6, recent work such as Rollout‑Marginal Distillation has also targeted long‑horizon generation; ID‑Forcing adds a complementary, training‑free option that could accelerate progress across the field.
Sources
Back to AIPULSEN