Explicit point correspondence is an attractive signal for adapting video generators: it tells the model how visible surfaces move from frame to frame. The catch, as the AcrossVAM2.0 paper notes, is that the same signal can overwrite the appearance prior that a large pretrained model has already learned. The authors frame this as a tension between motion supervision and appearance preservation.
AcrossVAM2.0 proposes conflict-aware point supervision to address that tension. The approach is designed for parameter-efficient adaptation, meaning it aims to adjust a pretrained video diffusion model without retraining everything. The "conflict-aware" part of the name signals that the method tries to recognise where point-correspondence supervision and the appearance prior are at odds, rather than applying the supervision uniformly.
The source is a single arXiv abstract, and it is truncated, so the specific mechanism and experimental results are not available here. No independent source is available to compare against; the summary above reflects only what the abstract states.