【必見！】進化版のAnimeteDiffが一気にレベルアップしたので紹介します！ 【stable diffusion】

Alice from AI’s in Wonderland introduces the AnimeteDiff extension, which allows for the creation of videos from Stable Diffusion images using text prompts. The video showcases the process without image adjustments, only through prompt input and settings on the Stable Diffusion WEB UI. The extension is capable of generating 2-second videos and has recently been updated to allow users to specify starting and ending images through the Control Net, enabling the linking of multiple short video clips. The video quality has been improved by TDS, and the process, while requiring some programming knowledge, is made accessible through Alice's guidance. The future of AI video creation is discussed, along with the current requirements for GPU memory and the potential for future updates to streamline the process.

The video script provides a step-by-step guide to installing the AnimeteDiff extension and the necessary motion modules. It emphasizes downloading modules from Google Drive to avoid issues and details the process of installing the extension through the Stable Diffusion WEB UI's Extensions page. The script also covers downloading and installing motion modules, potential issues with xformers, and the setup process for using AnimateDiff, including selecting the motion module, setting the number of frames and frames per second, and enabling the AnimateDiff feature. The use of the Mistoon Anime model, suitable for AnimateDiff, is highlighted, along with the prompt and settings used to generate the video.

The paragraph discusses the improvements made by TDS to enhance the video quality generated by the Stable Diffusion Web UI. It explains the incorporation of 'alphas cumprod' from the original repository into the DDIM schedule of the Web UI to achieve clearer images. The process involves downloading a JSON file called 'new schedule' and adding additional code to the DDIM.py file. The video demonstrates the significant difference in image quality before and after these enhancements, showcasing the potential for high-quality AI-generated videos.

Alice explains the installation and use of a control net to specify the starting and ending images of a video, allowing for greater control over the video's narrative. The script details the process of installing a specific branch of the control net from TDS's repository and replacing the existing hook.py file with a new one. It then demonstrates how to generate base images for the video frames using a specific model and prompt, and how to use the control net to control the start and end of the video. The video showcases the successful creation of a video with a controlled narrative and improved visual appeal.

The final paragraph of the script focuses on using a LoRA (Low-Rank Adaptation) to add special effects to the generated images and videos. Alice uses the Dragon Ball Energy Charge LoRA to create an image with an energy aura, which is intended to be used as the last frame of the video. She then creates the first frame with a similar composition and pose, using the control net to maintain consistency. The video is generated using AnimateDiff with specific settings, and Alice shares her excitement about the potential of this feature for creating high-quality, creative videos. She concludes by expressing her anticipation for the future development of AnimateDiff and ControlNet technologies.