Awesome-Video-Diffusion-Models

Organizes video diffusion research into a three-pillar taxonomy (video generation, video editing, video understanding) using a hub-and-spoke model where the survey document serves as the central organizing principle. The taxonomy implements nested subcategories (e.g., Text-to-Video subdivided into Training-based and Training-free approaches) with structured tables that systematically link to external papers, GitHub repositories, and project websites, enabling researchers to navigate the research landscape through semantic categorization rather than chronological or alphabetical ordering.

Provides structured comparison of text-to-video generation approaches by categorizing them into training-based methods (e.g., Make-A-Video, CogVideoX) and training-free methods, with linked papers and implementations for each. The capability enables researchers to understand the trade-offs between approaches that require fine-tuning on video datasets versus those that leverage pre-trained image diffusion models without additional training, facilitating architectural decision-making for practitioners building text-to-video systems.

Maintains bidirectional cross-references between research papers and their implementations, enabling practitioners to navigate from a paper to its GitHub repository and vice versa. The capability uses structured table entries that link papers (with arXiv/conference links) to corresponding GitHub repositories and project websites, creating a unified view of research and its practical instantiation. This supports practitioners who want to understand both the theoretical approach and the implementation details.

Provides citation information and academic usage guidance for the survey paper itself, enabling researchers to properly cite the comprehensive video diffusion survey in their own work. The capability includes BibTeX entries, citation formats, and information about the paper's publication in ACM Computing Surveys (CSUR), supporting academic reproducibility and proper attribution. This enables the survey to be used as an authoritative reference in academic work.

Organizes conditional video generation methods into pose-guided, motion-guided, sound-guided, and multi-modal control subcategories, with linked papers and implementations for each. The taxonomy enables practitioners to identify which conditioning modality (skeletal pose, motion vectors, audio, or combined inputs) best fits their use case, and to discover methods like AnimateAnyone and FollowYourPose that implement specific conditioning approaches. This capability maps user intents (e.g., 'animate a character from a pose sequence') to specific research papers and implementations.

Catalogs image-to-video (I2V) synthesis and animation methods with links to papers and implementations like Stable Video Diffusion and DynamiCrafter. The capability enables practitioners to discover methods that generate video sequences from static images, with subcategories distinguishing between pure I2V synthesis (generating motion from a single image) and animation approaches (bringing static artwork or illustrations to life). This supports use cases like creating video from photographs or animating artwork.

Organizes text-guided video editing methods into a structured catalog with links to papers and implementations that enable users to modify videos using natural language descriptions. The capability maps text prompts to video editing operations (e.g., 'change the sky to sunset', 'make the character smile'), enabling practitioners to discover methods that support semantic video manipulation without frame-by-frame manual editing. This differs from video generation by operating on existing video content rather than creating from scratch.

Catalogs multi-modal video editing methods that combine multiple input modalities (text, images, sketches, masks) to enable fine-grained control over video editing. The capability links to methods that support combined conditioning signals, enabling practitioners to discover approaches that go beyond text-only editing to incorporate visual constraints, spatial masks, or reference images. This supports complex editing workflows where text descriptions alone are insufficient.

Provides a structured index of video understanding and analysis research methods, enabling practitioners to discover approaches for video classification, action recognition, temporal reasoning, and semantic understanding. The capability catalogs papers and implementations that analyze video content rather than generate or edit it, supporting use cases like video captioning, action detection, and scene understanding. This represents the third pillar of the survey alongside generation and editing.

Catalogs datasets and evaluation metrics used in video diffusion research, enabling practitioners to understand how video generation, editing, and understanding methods are evaluated. The capability provides links to benchmark datasets (e.g., UCF101, Kinetics) and evaluation metrics (e.g., FVD, LPIPS, temporal consistency measures) used across the field, supporting practitioners in selecting appropriate evaluation approaches for their own systems. This enables informed comparison of methods and reproducible evaluation.

Implements a hub-and-spoke architecture that connects the survey to external resources including academic papers, GitHub repositories, project websites, and commercial platforms. The capability uses structured link patterns in README.md tables to systematically reference external implementations and research, creating a distributed knowledge network where the survey serves as the organizing principle while actual code and papers reside in external repositories. This enables practitioners to navigate from research concepts to implementations without leaving the survey context.

Curates a collection of visual demonstrations (GIFs, video clips) that illustrate key concepts and capabilities in video diffusion research. The capability organizes visual assets by type (algorithm demonstrations, motion examples, generation results, comparative examples) to provide practitioners with concrete examples of what different methods produce. This supports learning and evaluation by showing actual outputs rather than relying solely on text descriptions and paper figures.