ControlNetを使った「もっと簡単な」アウトペインティング#ai画像生成 #stablediffusion #ai技術

Alice, the assistant, introduces a new method of outpainting using ControlNet, which is easier than the previous Image-to-Image method. She plans to create a realistic image of a cosplayer dressed as Night from the Spy Family, specifically focusing on the character Lola in a black dress with cherry blossoms in the background. The model used will be Majimix Realistic V5, and the sample method will be DPM+2M. Despite potential compatibility issues with the model, Alice decides to proceed with it. She starts by increasing the batch size and using random seeds to explore different compositions, aiming to find a good layout for the painting.

Alice refines the image by adjusting the denoising strength using the XYZ plot, testing values between 0.3 and 0.7. She generates five images and decides that a denoising strength of 0.5 offers a good balance. She then experiments with the CFG scale, adjusting it from 5 to 11 to see how it affects the composition. After generating seven images, she finds that a CFG scale of 8.0 produces the most appealing expression. With the composition finalized, she closes the XYZ plot and sets the CFG scale to 8 to generate a single image for outpainting.

Alice discusses different resize modes for outpainting, comparing 'Just Resize', 'Crop & Resize', and 'Resize & Fill'. She explains the effects of each mode on the image, noting that 'Resize & Fill' is the most suitable for outpainting as it fills in the missing parts of the image. She demonstrates how to stretch a shield in the image by adjusting the size and using 'Resize & Fill'. Alice also shows how to upscale the image while maintaining the aspect ratio and discusses the importance of doing so gradually to preserve the image's integrity.

Alice makes final adjustments to the image, focusing on areas that need improvement such as the hand and the cherry blossoms. She uses the ControlNet to make these adjustments and shares tips for using the 'Just Resize' mode and the Aspect Ratio Selector extension for creating images with specific aspect ratios. She also addresses potential issues with image integration when upscaling and suggests using a stronger denoising strength for more natural results. Alice concludes by discussing the time-consuming nature of image generation and advises viewers to be mindful of their VRAM capacity when generating large images.