APLA: Additional Perturbation for Latent Noise with Adversarial Training Enables Consistency

arXiv:2308.12605v2 Announce Type: replace
Abstract: Diffusion models have exhibited promising progress in video generation. However, they often struggle to retain consistent details within local regions across frames. One underlying cause is that traditional diffusion models approximate Gaussian noise distribution by utilizing predictive noise, without fully accounting for the impact of inherent information within the input itself. Additionally, these models emphasize the distinction between predictions and references, neglecting information intrinsic to the videos. To address this limitation, inspired by the self-attention …

abstract accounting adversarial adversarial training arxiv consistent cs.ai cs.cv diffusion diffusion models distribution however impact noise predictive progress struggle training type video video generation