【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】
TLDRIn this informative video, the creators from AI's Wonderland, Alice and Yuki, guide viewers on how to use Supermerger to craft a personalized merge model akin to majicmix, which can be utilized through CIVITAI. They emphasize the importance of understanding licensing terms before distribution, and suggest that for now, one should focus on merging models for personal use. The video delves into the intricacies of merging, including selecting models, choosing a merge mode, and adjusting alpha and beta values to fine-tune the merge ratio. It also touches on the legal considerations of different models and licenses, such as CreativeML Open RAIL-M and Deliberate's Attribution-nonCommercial-NoDerive restrictions. The hosts demonstrate the installation process of Supermerger and its user interface, showcasing how to merge models and generate images without creating a new file each time, thus saving storage space. They also introduce advanced features like hierarchy merging, which allows for the adjustment of the merging ratio at each U-Net layer, and the use of presets for different effects. The video concludes with a discussion on saving the merged model and the various calculation modes available in Supermerger, encouraging viewers to experiment and create their own unique models.
Takeaways
- 📝 **License Consideration**: When creating and potentially distributing a merge model, it's crucial to understand and adhere to the licensing terms to avoid ethical and legal issues.
- 🤖 **Model Merging**: Supermerger allows users to combine different models, such as Majicmixrealistic and epiCRealism, to create a new model without the need for extensive technical expertise.
- 🔄 **Merge Modes**: Users can select different merge modes like Weight sum to control how models are combined, with alpha values determining the influence of each model in the merge.
- 💾 **Storage Efficiency**: Supermerger enables on-the-fly model merging without creating large checkpoint files each time, saving storage space.
- 🚀 **Instant Model Generation**: After merging, users can immediately generate images to see the results, offering a quick way to visualize the effects of the merge.
- 📈 **Adjustable Parameters**: The merging process includes adjustable parameters such as alpha and beta values, and the ability to select the calculation mode for fine-tuning the merge.
- 📊 **XYZ Plot Comparison**: Supermerger provides an XYZ plot feature for comparing different merge ratios and model generations side by side.
- 🧠 **Hierarchy Merging**: A more advanced feature is hierarchy merging, which allows users to adjust the merge ratio for each U-Net layer, influencing different aspects of the generated image.
- 🎰 **Random Merging**: For a chance-based approach, users can opt for random merging, which applies random values to the hierarchy merging process.
- 🔗 **Resource References**: The script references external resources like Tofu no Kakera's page for more detailed explanations on certain merging techniques and calculation modes.
- 🔖 **Saving Custom Models**: Once a satisfactory merge is achieved, users can save their custom model with a chosen name, making it easier to reuse and share (while respecting licensing).
Q & A
What is the main topic of the video?
-The main topic of the video is about using Supermerger to create a personal merge model, similar to majicmix, and the considerations regarding licensing when distributing such models.
Who are the speakers in the video?
-The speakers in the video are Alice from AI’s, in Wonderland and Yuki.
What is the importance of checking the license when distributing a merge model?
-Checking the license is important to ensure that the distribution and use of the merge model comply with the terms set by the original creators, avoiding any legal or ethical issues.
What does the term 'Weight sum' refer to in the context of merging models?
-In the context of merging models, 'Weight sum' is a merge mode that allows for the combination of two models by assigning a specific weight (alpha value) to each, determining their influence on the final merged model.
How can one save the merged model as a file?
-To save the merged model as a file, one needs to open the Save Settings tab, check the 'save model' option, and provide a custom name for the merged model.
What is the purpose of the 'Hierarchy merging' feature in Supermerger?
-The 'Hierarchy merging' feature allows users to change the merging ratio for each U-Net layer, enabling more control over which aspects of the models are combined and how they influence the final image.
What does the video suggest regarding the distribution of models through CIVITAI?
-The video suggests that while it is possible to distribute models through CIVITAI, one must be cautious and ensure they understand and comply with the licensing terms of the models they are using.
What is the significance of the 'Random mode' in hierarchy merging?
-The 'Random mode' in hierarchy merging is used to perform random merges at various ratios, which can quickly generate a variety of images and is useful for exploring different outcomes without manually setting each layer's merge ratio.
How does the video address the issue of storage space when merging models?
-The video highlights that SuperMerger allows users to merge models without creating a new file each time, thus reducing the pressure on storage space as large checkpoint files do not need to be saved after every merge.
What is the role of the 'XYZ plot' tab in comparing merge models?
-The 'XYZ plot' tab is used to visually compare different merge models by generating a grid of images with varying alpha values, seed values, and prompts, allowing users to see the effects of different merge ratios and settings.
What are some of the presets available in SuperMerger for hierarchical merging?
-Some of the presets available in SuperMerger for hierarchical merging include GRAD_V, GRAD_A, WRAP12, RING10 3, and others that allow for different combinations and influences of the models being merged.
Outlines
🚀 Introduction to Supermerger and Licensing Considerations
Alice and Yuki introduce the concept of using Supermerger to create a custom merge model similar to Majicmix, emphasizing the importance of licensing when considering distribution. They discuss the different types of models available on CIVITAI and the need to understand the license agreements before distributing or selling products using these models. The video also touches on the complexity of licenses and suggests that viewers consult the license texts for models like CreativeML Open RAIL-M and stable diffusion. The presenters also mention the prohibitions associated with certain models and recommend choosing a model listed on CIVITAI for those concerned about licensing issues.
🛠️ SuperMerger Installation and Basic Usage
The video provides a step-by-step guide on installing Supermerger as an extension to stable diffusion web ui. It explains how to access the SuperMerger tab and merge two models, Model A and Model B, using the Weight sum mode. The presenters demonstrate how to adjust the alpha value to control the influence of Model B in the merged model and generate an image using the Merge & Gen button. They also discuss how SuperMerger allows users to generate images without saving a merge model file, thus saving storage space, and how to save a merge model if desired.
📈 Advanced Merging Techniques and History Management
The video delves into advanced merging techniques, including hierarchy merging and the use of presets. It explains how to adjust the merging ratio for each U-Net layer and how different layers affect various parts of the generated image. The presenters also discuss how to use the History tab to recall previously merged models and the XYZ plot for comparing different merge ratios. They highlight the convenience of the Save Settings tab for labeling and recalling images based on their merge model ID.
🎨 Customizing Merging Ratios and Exploring Effects
The video explores manual adjustment of merging ratios and presets for hierarchy merging. It presents different presets like GRAD_V and GRAD_A and demonstrates their effects on the generated image. The presenters also discuss the impact of different layers on the image and how to use the XYZ plot for comparing images generated with various calculation modes. They provide insights into the characteristics of each preset and how they affect the final output.
🔄 Random Merging and Model Saving
The video introduces the concept of random merging using the 'let the dice roll' method, which allows for a variety of hierarchical merges and quick generation of diverse images. It also covers the process of saving the merged model with a custom name and as a safetensors file. The presenters encourage viewers to experiment with different merging methods and calculation formulas, and they provide resources for further learning.
📺 Conclusion and Engagement Invitation
The video concludes with a call to action for viewers to subscribe to the channel and like the video. The presenters express gratitude for watching the video to the end and bid farewell, promising to see the audience in the next video.
Mindmap
Keywords
💡Supermerger
💡Merge Model
💡CIVITAI
💡License
💡Stable Diffusion
💡EpiCRealism
💡Weight Sum
💡Hierarchy Merging
💡XYZ Plot
💡Random Merge
💡Calculation Mode
Highlights
Alice and Yuki from AI’s, in Wonderland introduce the use of Supermerger to create a personalized merge model.
The importance of considering licensing when distributing models through platforms like CIVITAI is emphasized.
Different models, such as epiCRealism and Majicmix, can be merged to create a unique model without initial licensing issues.
The stable diffusion model's license allows for distribution and sale of products as long as they are used ethically.
Deliberate's license, Attribution-nonCommercial-NoDerive, prohibits commercial use and sharing of merge models.
Supermerger allows users to merge models without creating a large file each time, saving storage space.
The merging process can be visualized and compared using the XYZ plot feature in Supermerger.
Hierarchy merging in Supermerger enables users to change the merging ratio for each U-Net layer, affecting different parts of the generated image.
Presets in Supermerger provide various merging characteristics that can be easily selected and applied.
The 'let the dice roll' feature in Supermerger offers a random merge approach for a quick variety of images.
Different calculation modes in Supermerger can produce varying results in image generation.
SuperMerger supports merging three models and offers various merging methods with listed calculation formulas.
The video provides a step-by-step guide on installing Supermerger as an extension for stable diffusion web ui.
The Weight sum merge mode in Supermerger uses an alpha value to determine the influence of Model B in the merged model.
The generated merge model can be saved as a file with a custom name in the stable diffusion webui folder.
The History tab in Supermerger allows users to load previously merged models and their settings.
The merge model's ID can be saved into the generated image for easy recall and comparison.
The video concludes with a demonstration of creating an original model using Supermerger and the various features explored.