**Video Interpolation with Diffusion Models** **Authors:** Siddhant Jain **Abstract:** This paper introduces VIDIM, a generative model for video interpolation that creates short videos given start and end frames. VIDIM uses cascaded diffusion models to first generate the target video at a low resolution and then refine it to a high resolution. The model is designed to handle complex, nonlinear, and ambiguous motions, which are challenging for previous state-of-the-art methods. VIDIM achieves high-fidelity results by conditioning the super-resolution model on the original high-resolution frames and using classifier-free guidance on the start and end frames. The model is fast, requires less than a billion parameters, and scales well with larger parameter counts. **Introduction:** Diffusion models have gained popularity for generative tasks due to their stability and ability to produce high-quality samples. Video interpolation, a specific task within this domain, aims to generate intermediate frames between two consecutive frames. Prior methods often fail when the motion between the start and end frames is complex or ambiguous. VIDIM addresses these limitations by explicitly conditioning the generative model on the start and end frames, allowing it to produce more plausible and high-quality results. **Methodology:** VIDIM consists of a base model and a super-resolution model. The base model generates 7 low-resolution frames between the start and end frames, while the super-resolution model refines these frames to high resolution. The models are trained using a cascaded approach, where the base model is conditioned on the start and end frames, and the super-resolution model is conditioned on the low-resolution frames and the original high-resolution frames. This setup ensures that the model can learn to handle complex motions and produce sharp, natural-looking results. **Experiments:** VIDIM is evaluated on the Davis-7 and UCF101-7 datasets, which contain challenging examples with large and ambiguous motions. The model is compared against several baselines, including RIFE, FILM, LDMVFI, and AMT. VIDIM consistently outperforms these methods in terms of reconstruction metrics (PSNR, SSIM, LPIPS) and generative metrics (FID, FVD). Human evaluation further confirms the superior quality of VIDIM samples. **Discussion and Future Work:** The paper discusses the importance of explicit frame conditioning and classifier-free guidance in achieving high-quality results. Future work could explore applications of VIDIM in frame expansion, video restoration, and other video generation tasks. The scalability of VIDIM is also demonstrated, showing that it can handle larger parameter counts and higher resolutions without significant degradation in quality. **References:** The paper cites relevant literature on diffusion models, video interpolation, and related techniques, providing a comprehensive background for the proposed method.**Video Interpolation with Diffusion Models** **Authors:** Siddhant Jain **Abstract:** This paper introduces VIDIM, a generative model for video interpolation that creates short videos given start and end frames. VIDIM uses cascaded diffusion models to first generate the target video at a low resolution and then refine it to a high resolution. The model is designed to handle complex, nonlinear, and ambiguous motions, which are challenging for previous state-of-the-art methods. VIDIM achieves high-fidelity results by conditioning the super-resolution model on the original high-resolution frames and using classifier-free guidance on the start and end frames. The model is fast, requires less than a billion parameters, and scales well with larger parameter counts. **Introduction:** Diffusion models have gained popularity for generative tasks due to their stability and ability to produce high-quality samples. Video interpolation, a specific task within this domain, aims to generate intermediate frames between two consecutive frames. Prior methods often fail when the motion between the start and end frames is complex or ambiguous. VIDIM addresses these limitations by explicitly conditioning the generative model on the start and end frames, allowing it to produce more plausible and high-quality results. **Methodology:** VIDIM consists of a base model and a super-resolution model. The base model generates 7 low-resolution frames between the start and end frames, while the super-resolution model refines these frames to high resolution. The models are trained using a cascaded approach, where the base model is conditioned on the start and end frames, and the super-resolution model is conditioned on the low-resolution frames and the original high-resolution frames. This setup ensures that the model can learn to handle complex motions and produce sharp, natural-looking results. **Experiments:** VIDIM is evaluated on the Davis-7 and UCF101-7 datasets, which contain challenging examples with large and ambiguous motions. The model is compared against several baselines, including RIFE, FILM, LDMVFI, and AMT. VIDIM consistently outperforms these methods in terms of reconstruction metrics (PSNR, SSIM, LPIPS) and generative metrics (FID, FVD). Human evaluation further confirms the superior quality of VIDIM samples. **Discussion and Future Work:** The paper discusses the importance of explicit frame conditioning and classifier-free guidance in achieving high-quality results. Future work could explore applications of VIDIM in frame expansion, video restoration, and other video generation tasks. The scalability of VIDIM is also demonstrated, showing that it can handle larger parameter counts and higher resolutions without significant degradation in quality. **References:** The paper cites relevant literature on diffusion models, video interpolation, and related techniques, providing a comprehensive background for the proposed method.