AI繪畫Stable Diffusion,幾種手指修正方法

The speaker begins by addressing the common issue of incorrect finger rendering in images generated by stable diffusion. They use an example image with an extra two fingers and suggest searching for similar finger images on Google or taking a photo of one's own hand. Various types of hand gesture images, including real and cartoon images, can be used for reference. The speaker demonstrates how to find a suitable hand image, adjust its size and position in Photoshop (PS), and use the ControlNet Canny option to correct the fingers. The process involves adjusting the image's transparency for easier editing and using the ControlNet with specific weight and guidance start values to achieve the desired result.

In this paragraph, the speaker discusses the use of a script labeled 'XYZ' to test different ControlNet weight and guidance start values to find the best settings for correcting finger issues. They mention that the tested images, along with the corresponding XY data, can be found in a folder named 'text to image grids'. The speaker emphasizes the importance of testing different values to achieve the most natural-looking fingers. They also share that the test images will be uploaded to a cloud storage service for reference. The speaker then demonstrates how to input the effective values into the ControlNet settings in stable diffusion and mentions that typical values for the weight are around one, and for guidance start, they are often between 0.1 and 0.2.

The speaker continues by describing a second method to fix finger issues using the Canny option in ControlNet. They save the previously corrected hand image and open it in Photoshop to add missing details using the pencil tool. The speaker explains how to adjust the brush size and switch between black and white colors to clean up unwanted finger lines. After making the necessary adjustments, they save the image and return to stable diffusion to apply the changes using the ControlNet with the Canny model. They also mention using the 'XYZ' script to test different weight and guidance start values for further refinement.

In this section, the speaker introduces a third method for correcting finger issues using the SD WebUI Openpose Editor extension. They explain how to drag a problematic image into the ControlNet and select the Openpose preprocessor. Once the preprocessor is run, the speaker demonstrates how to edit the character's pose and hand gestures using various tools provided by the editor. They discuss the process of hiding parts of the image, mirroring, and adjusting the size and position of the hand to correct the fingers. The speaker also mentions the use of the 'XYZ' script to find the best parameters for finger correction.

The speaker further elaborates on the use of Openpose for full-body image correction, including the hands. They discuss the process of adjusting the Openpose model to fit the image without fingers and then saving the image for use in the depth map editor extension. The speaker also mentions that if no suitable hand gestures are found, one can start with an Openpose model without fingers and adjust it accordingly. They provide a demonstration of how to add a background image and a hand gesture to the image using the depth map editor and send it to ControlNet for further processing.

The speaker provides guidance on installing the depth map editor extension, which is necessary for correcting finger issues in images. They mention downloading a specific package from civitai.com and placing it in the appropriate folder within the stable diffusion webUI. The speaker also addresses potential errors that may occur during installation and provides a solution by suggesting modifications to the 'serializing.py' file. They describe the process of adding background images and desired poses using the 3D Openpose extension and adjusting the character's position and size within the image.

In the final paragraph, the speaker discusses the use of 3D Openpose for adjusting hand gestures and mentions that the test images will be shared on a cloud storage service. They provide a brief overview of the parameters used for creating new hand gestures and suggest that viewers refer to the shared images for guidance. The speaker concludes by encouraging viewers to follow for updates and thanks them for watching the video.