Refines prompts with parallel or sequential expert agents to boost cultural fidelity in text-to-video generation across mono- and cross-cultural prompts.

github.com/AIM-SCU/CRAFT

arxiv.org/abs/2605.16716

Decouples fairness from guidance scale in diffusion models by equalizing guidance distributions or offsetting null embeddings

github.com/Kim-Myeong-Soo/stay-fair

arxiv.org/abs/2605.28036

BiDPO jointly optimizes image and text preferences with region-guided alignment to boost compositional fidelity in T2I generation.

arxiv.org/abs/2605.28615

github.com/anzeameol/BiDPO

Boosts text-to-video efficiency by combining hybrid full-sparse attention with Skiparse-2D and Sparse Sequence Parallelism to cut communication and maintain quality.

arxiv.org/abs/2605.28691

Directs narrative-aware video generation by conditioning on multiple keyframes, enabling single-shot, multi-shot, and extension scenarios

arxiv.org/abs/2605.27891

Unifies layers in a 20B masked region diffusion model for multi-layer transparent image generation and editing

arxiv.org/abs/2605.27235

Enables temporally coherent video generation with decentralized training, doubling FVD improvement and boosting CLIP similarity and aesthetics under equal compute.

arxiv.org/abs/2605.26064