SDXL真人LoRa訓練教學，從素材到模型一次告訴你！

SDXL真人LoRa訓練教學，從素材到模型一次告訴你！

Table of Contents

1. Introduction
2. Benefits of SDXL Training in Realism
3. The Importance of DeepdanBooru for Anime Models
4. Collecting Materials for SDXL Training
5. Tagging Images for SDXL Training
6. Efficient Tag Review and Removal
7. Choosing the Right Base Model for Training
8. Preparing the Training Folder
9. Setting Training Parameters for Different VRAM
   • 9.1 VRAM Configuration for 8G
   • 9.2 VRAM Configuration for 12G
   • 9.3 VRAM Configuration for 24G
10. Fine-tuning Parameters for Training
11. Using XYZ Charts to Select Models
12. Conclusion

Introduction

SDXL is a powerful tool for training realistic models. In this article, we will discuss important concepts related to SDXL training, from VRAM settings to material selection and model choices. We will also explore the community's experiences and provide detailed instructions to help You optimize your training process. By the end of this article, you will have a comprehensive understanding of SDXL training and be able to streamline your training Journey.

Benefits of SDXL Training in Realism

SDXL's training capabilities in the realm of realism are truly impressive. Compared to previous methods, SDXL allows for more natural Prompts and reduces the need for excessive text inputs. When it comes to training real-world models, SDXL excels in capturing nuances and producing more accurate results. However, it is important to note that promoting the development of anime models still requires community involvement.

The Importance of DeepdanBooru for Anime Models

DeepdanBooru has been influential in the auto-tagging of anime models and is the predecessor of SDXL's tagging system. However, it is essential to have a high-caliber deepdanBooru model specifically designed for anime models. While SDXL makes training easier and more effortless, it is crucial to have a reliable deepdanBooru model. Although SDXL does not rely on deepdanBooru for tagging, it is still recommended to utilize the tagger extension for accurate tagging.

Collecting Materials for SDXL Training

When collecting materials for SDXL training, it is advisable to Gather around 20 images of your desired subject, prioritizing half-body or headshots. Including a few full-body images is also beneficial. It is important to ensure minimal age gap between the images to avoid inconsistencies. As for image Dimensions, SDXL does not have strict requirements and accepts various ratios. Once the materials are gathered, it is crucial to tag them correctly. While manual labeling is time-consuming, using the tagger extension with deepdanBooru tags can yield accurate results.

Efficient Tag Review and Removal

Reviewing and removing tags efficiently is crucial for quality training. One recommended tool for this purpose is the BooruDatasetTagManager, which provides easy tag inspection and removal. By loading your material folder and enabling translations to Traditional Chinese, you can effectively manage tags. Unnecessary tags, particularly those that capture the appearance features of the subject, should be removed to facilitate accurate training. Retaining only background-related tags is recommended to allow for variability in clothing and backgrounds.

Choosing the Right Base Model for Training

Selecting the right base model for training greatly influences the success of your results. The key is to choose a model that is closest to your target subject. For instance, SD1.5 real-life models excel in drawing beautiful women, while special cases like male celebrities may benefit from fine-tuning an Asian male model. By reducing the gap between the base model and the desired outcome, the chances of overfitting also decrease.

Preparing the Training Folder

To begin the training process, utilize the Folder Preparation function in Kohya GUI. This function automatically organizes the training data folders according to Kohya's naming conventions. While the tool is no longer actively maintained, a backup link will be provided if it becomes unavailable. In this step, you will need to specify instance prompt (trigger keyword) and class prompt (category keyword). These prompts guide the model in recognizing and generating the desired output. Ensure your training images are properly sourced and set the appropriate repeat value, depending on the number of desired training cycles. Regularization images can be excluded, and the destination training directory should be specified.

Setting Training Parameters for Different VRAM

Optimizing training parameters Based on different VRAM sizes is essential to achieve optimal results. The following configurations are recommended:

• For 8G VRAM: Use 4090 configuration with 1 batch size, cache latent, and disk settings. Set the maximum resolution to 512x512.
• For 12G VRAM: Utilize a similar configuration as 8G but increase the batch size to 6 and set the maximum resolution to 768x768.
• For 24G VRAM: Adjust the configuration for 12G VRAM by further increasing the batch size to 8 and setting the maximum resolution to 1024x1024. Disable the text encoder outputs caching.

Fine-tuning Parameters for Training

Fine-tuning training parameters is crucial for obtaining the desired results. While there is no definitive best setting, the following parameters have proven effective: using the SDXL Loha model, constant scheduler, no warm-up, and an additional set of G-recommendation optimizer parameters. It is important to note that these parameters are not mandatory and may require personal experimentation.

Using XYZ Charts to Select Models

XYZ charts provide a comprehensive evaluation of models from various perspectives. By analyzing different models based on their composition, resemblance, and imitation of specific features, you can make more informed decisions. However, it is important to remember that XYZ charts alone may not be sufficient. The community is actively working on developing new methods for evaluating models, and user feedback and participation are highly encouraged.

Conclusion

SDXL training offers immense potential for creating highly realistic models. By understanding the key concepts and following the recommended procedures outlined in this article, you can streamline your training process and achieve optimal results. Remember to adapt the training parameters to your specific VRAM size and fine-tune the settings based on your desired outcomes. Additionally, XYZ charts can aid in selecting the most suitable models. Be sure to actively participate in the community to further advance SDXL training techniques.

Highlights

• SDXL training allows for more natural prompts and reduces text inputs.
• DeepdanBooru is crucial for auto-tagging anime models in SDXL.
• Collect around 20 images of your subject, prioritize half-body or headshots.
• Use the BooruDatasetTagManager tool for efficient tag review and removal.
• Choose a base model closest to your target subject for successful training.
• Utilize Folder Preparation in Kohya GUI to organize the training data.
• Adjust training parameters based on your VRAM size for optimal results.
• Fine-tune parameters and use XYZ charts for model selection.
• Active community participation in evaluating new training methods is encouraged.

FAQ

Q: Can SDXL be used to train models for anime characters? A: Yes, SDXL is highly effective in training models for anime characters. However, it is important to have a reliable deepdanBooru model specifically designed for anime tagging.

Q: How many images should I collect for training in SDXL? A: Collecting around 20 images of your subject is recommended for SDXL training. Prioritize half-body or headshots and ensure minimal age gap between images.

Q: Is it necessary to manually label images in SDXL? A: While manual labeling is time-consuming, you can still use the tagger extension with deepdanBooru tags for accurate labeling. However, it is important to review and remove unnecessary tags.

Q: How should I choose the base model for training in SDXL? A: Select a base model that is closest to your target subject. SD1.5 models excel in drawing beautiful women, but for specific cases like male celebrities, fine-tuning an Asian male model may be more suitable.

Q: How can I optimize training parameters for different VRAM sizes? A: Adjusting VRAM-specific configurations such as batch size, maximum resolution, and caching settings will optimize training parameters for different VRAM sizes.

Q: What tools can I use to efficiently review and remove tags in SDXL? A: The BooruDatasetTagManager tool facilitates efficient tag review and removal in SDXL.