Rectified SpaAttn: Revisiting Attention Sparsity for Efficient Video Generation

A new approach called Terminal Velocity Matching (TVM) has been proposed, which generalizes flow matching to enhance one- and few-step generative modeling. TVM focuses on the transition between diffusion timesteps and regularizes behavior at terminal time, proving to provide an upper bound on the 2-Wasserstein distance between data and model distributions under certain conditions.

A new approach called Cross-Resolution Phase-Aligned Attention (CRPA) has been introduced to address a critical failure in the use of rotary positional embeddings (RoPE) within Diffusion Transformers, particularly when handling mixed-resolution denoising. This issue arises from linear interpolation that leads to phase aliasing, causing instability in the attention mechanism and resulting in artifacts or collapse during processing.

The introduction of MapReduce LoRA and Reward-aware Token Embedding (RaTE) marks a significant advancement in optimizing generative models by addressing the alignment tax associated with multi-preference optimization. These methods enhance the training of preference-specific models and improve token embeddings for better control over generative outputs. Experimental results demonstrate substantial performance improvements in both text-to-image and text-to-video generation tasks.

A new framework named Plan-X has been introduced to enhance video generation through high-level semantic planning, addressing the limitations of existing Diffusion Transformers in visual synthesis. The framework incorporates a Semantic Planner, which utilizes multimodal language processing to interpret user intent and generate structured spatio-temporal semantic tokens for video creation.

A new study has introduced a framework for deterministic inference across varying tensor parallel sizes, addressing the issue of training-inference mismatch in large language models (LLMs). This mismatch arises from non-deterministic behaviors in existing LLM serving frameworks, particularly in reinforcement learning settings where different configurations can yield inconsistent outputs.