TL;DRGenerates videos of a user-specified subject with no per-subject tuning by separating identity injection (from subject images) and motion learning (from unrelated videos), matching prior zero-shot methods at roughly 1% of the compute.

TL;DRA large open dataset for reference-image-guided editing, built via "reference completion" — automatically synthesizing the missing reference for existing text-edit triplets to yield ~477K input–reference–instruction–target quadruplets across six edit categories.

TL;DRReconstructs dynamic 3D humans and their surrounding scene over time from only a few low-overlap cameras, using camera-controlled video diffusion to synthesize dense novel views and reconstructing background and people separately with cross-view tracking.

TL;DRTackles open-domain customized video generation — synthesizing videos of user-specified subjects across arbitrary domains — with an end-to-end ecosystem spanning data, model, and evaluation.

Qualitative comparison · "a dog on top of a purple rug in a forest"

CustomDiT (rightmost) best preserves subject identity while generating natural, dynamic videos.

TL;DRIntroduces online event-boundary detection for streaming video (unlike prior methods that need the whole clip), with an Event-Segmentation-Theory–inspired model (ESTimator) that flags boundaries where its frame predictions diverge from what actually happens.

TL;DREditing methods built for U-Net diffusion fail on multimodal diffusion transformers (MM-DiT); this work analyzes MM-DiT's text–image attention to build an editing method handling global-to-local edits across MM-DiT variants, including few-step models.

TL;DRSupports long, multi-step image edits by storing past edits in a layer-wise memory, with background-consistency guidance and multi-query disentanglement so objects can be added or removed with rough masks while keeping earlier edits intact.

TL;DRReplaces costly 1×1 channel-squeeze layers in ConvNets by splitting channels into subsets and dynamically picking per pixel based on the input — about 25% faster at comparable accuracy, and reusable as a downscaler and dynamic pruner.

TL;DRGenerates coherent image sequences from multi-sentence stories using a context-memory module that propagates information across frames and an online text-augmentation scheme that creates pseudo-descriptions during training, improving long-term consistency.