Image-to-Video Transfer Learning based on Image-Language Foundation Models: A Comprehensive Survey

This project is an academic survey paper rather than a software product. While it provides a valuable taxonomy for researchers, it possesses no technical moat or proprietary code. The defensibility is low (2) because the value lies in the synthesis of existing research, which is trivially reproducible by any domain expert or even high-end LLMs today. The frontier risk is high because labs like OpenAI, Google (DeepMind), and Meta are moving beyond 'image-to-video transfer' (the hacky adaptation of 2D models to 3D temporal data) and toward native video foundation models (e.g., Sora, Veo, Movie Gen). The 7 forks against 0 stars suggest it is being used by a small group of researchers as a bibliography or reference list. From a competitive standpoint, this is a 'map' of a rapidly evolving territory; the map becomes obsolete as soon as the frontier labs release the next generation of native video-text models, rendering the complex 'adaptation' techniques summarized here redundant.