【絶対できる】Supermergerの階層マージを使いこなして、myマージモデルを作ろう【stable diffusion】

The video begins with an introduction to the use of Supermerger for creating custom merge models, similar to majicmix, and distribution through CIVITAI. The importance of checking licenses for distribution is emphasized, with a brief overview of the types of models available on CIVITAI, including epiCRealism and CreativeML Open RAIL-M. The video also touches on the licensing terms for stable diffusion and the potential for commercial use of model outputs. The complexity of understanding licenses is acknowledged, and viewers are encouraged to comment with feedback or corrections.

This section details the process of setting up Supermerger, which requires the stable diffusion web UI. Instructions for installation are provided, including copying and pasting the URL into the Extensions tab for installation. The video then moves on to demonstrate the creation of a merge model, with a focus on combining Japanese real and foreign real models to achieve a half-real, transcendental beauty. The use of Majicmixrealistic and epiCRealism models is highlighted, and the process of merging models using the Weight sum mode is explained, including the adjustment of alpha values and the generation of images.

The paragraph discusses the exploration of merge ratios using the XYZ plot feature in Supermerger. It explains how to compare different merge ratios by adjusting alpha values and generating images in a grid format. The use of random seed values for generating diverse images and the convenience of recalling previous merge models through the History tab are also covered. The video then delves into hierarchy merging, explaining the concept of U-Net layers and how adjusting the merging ratio at each layer can affect the final image. The process of using hierarchy merging is demonstrated, with a focus on the influence of different layers on the image's features.

This section continues the discussion on hierarchical merging, showcasing the impact of different layers on the image's appearance. The video presents various presets and their effects on the merge, such as GRAD_V and GRAD_A, and compares the results. It also explores the influence of the shallow and middle layers on facial features and skin texture. The video then examines the impact of OUT layers on the overall image and compares different merging strategies, including WRAP12 and RING10 3. The goal is to achieve a balance between the characteristics of both models used in the merge.

The video introduces the concept of random hierarchical merging, allowing users to explore a variety of merge ratios without manual adjustment. The 'let the dice roll' feature is highlighted for its ease of use and the ability to generate diverse images quickly. The video also discusses different calculation modes, such as smoothAdd and smoothAdd,MT, and their impact on image generation. The use of XYZ plots to compare calculation modes is demonstrated, and the video concludes with a recommendation to refer to Tofu no Kakera's detailed explanation of calculation modes for further understanding.

The video concludes with a summary of the key points covered, including the setup and use of Supermerger, the exploration of merge ratios, and the intricacies of hierarchical merging. The presenter encourages viewers to subscribe to the channel and like the video, and expresses satisfaction with the merge results achieved throughout the video. The video ends with a farewell and an anticipation for the next video installment.