ID-Forcing extends short-horizon video diffusion models to minute-scale generation
SeoulNatlUniv · hf · 2026-10-06
Autoregressive video diffusion models drift on long videos as colors/textures shift and motion decays. The authors identify the root cause: KV conditioning alone assumes cached entries stay in-distribution, but beyond the training horizon nothing constrains KV construction during rollout (the KV-provenance problem), so the cache itself goes OOD.
- ID-Forcing is a test-time framework aligning KV caching and conditioning with training configurations
- Key mechanism, self-caching: each chunk is cached without attending to prior KVs, keeping the rolling window exactly in-distribution
- Seamlessly extends short-horizon models to minute-scale video, outperforming prior work on drift (per new drift metrics and a user study) while staying competitive on standard benchmarks
More from Multimodal
- Seedance v2.5 prompt breakdown: a 60s vertical Korean drama short with hard cuts — creatoroff · 2026-10-06
- Feeding an AI-rendered persona into Suno yields a surprisingly good longtermist metal song — anderssandberg · 2026-10-06
- Identity LoRA Trained on 131 Images Still Drifts on Faces — Reddit Asks for Proven ComfyUI Workflow — DragonfruitNo9822 · 2026-10-06
- Small Team Makes 15-Min AI Episodes in Two Weeks for $15K Total — lmoroney · 2026-10-06
- Dark symphonic rock AI music video 'WHERE THE DNIPRO SLEEPS' — aky-an · 2026-10-06
- A CHICKEN DIES: A generative sound-design music arcade game where your moves are the drums — jordiponsdotme · 2026-10-06