Which Layer Causes Distribution Deviation? Entropy-Guided Adaptive Pruning for Diffusion and Flow Models

MoGAN has been introduced as a motion-centric post-training framework aimed at enhancing motion quality in video diffusion models, which often struggle with issues like jitter and ghosting. This framework utilizes a DiT-based optical-flow discriminator to improve motion realism without relying on reward models or human preference data.

The introduction of Filter Like You Test (FLYT) presents a novel algorithm for curating large-scale vision-language datasets, enhancing the selection of pretraining examples by learning the usefulness of each data point through gradient signals from downstream tasks. This is complemented by Mixing-FLYT (M-FLYT) and Soft Cap Sampling (SCS), which improve dataset filtering and accuracy.

A recent study on Vision Transformers (ViTs) reveals a non-monotonic scaling behavior, where deeper models like ViT-L may underperform compared to shallower variants such as ViT-S and ViT-B. This research identifies a three-phase pattern—Cliff-Plateau-Climb—indicating how representation quality evolves with depth, particularly noting the diminishing role of the [CLS] token in favor of patch tokens for better performance.

A novel approach called Low-Temperature Distillation (LTD) has been introduced to enhance adversarial training in neural networks, addressing the vulnerabilities associated with one-hot label representations in image classification. LTD utilizes a lower temperature in the teacher model while keeping the student model's temperature fixed, refining label representations and improving model robustness against adversarial attacks.

The introduction of DP-MicroAdam marks a significant advancement in the realm of adaptive optimizers for differentially private training, demonstrating superior performance and convergence rates compared to traditional methods like DP-SGD. This new algorithm is designed to be memory-efficient and sparsity-aware, addressing the challenges of extensive compute and hyperparameter tuning typically associated with differential privacy.

A recent study titled 'ModHiFi: Identifying High Fidelity predictive components for Model Modification' explores methods to modify open weight models without access to training data or loss functions. The research focuses on identifying critical components that influence predictive performance using only distributional access, such as synthetic data.

The introduction of Multi-Granularity Differentiable Architecture Search (MG-DARTS) marks a significant advancement in neural architecture search (NAS), focusing on optimizing both model effectiveness and efficiency. This framework addresses limitations in existing differentiable architecture search methods by incorporating finer-grained structures, enhancing the balance between model performance and size.

The newly proposed DeCo framework introduces a frequency-decoupled pixel diffusion method for end-to-end image generation, addressing the inefficiencies of existing models that combine high and low-frequency signal modeling within a single diffusion transformer. This innovation allows for improved training and inference speeds by separating the generation processes of high-frequency details and low-frequency semantics.