Create consistent characters with Stable diffusion!!
TLDRThe video script outlines a method for creating and training AI-generated characters with consistent appearances across different images. It details a three-part process involving character creation, refinement through editing and retouching, and finally, training a model using a dataset of the character in various poses and settings. The script emphasizes the importance of high-quality character sheets, the use of control nets and stable diffusion for image generation, and the meticulous cleanup and preparation of images for training. It also discusses the potential for infinite refinement through retraining with new generated images and invites viewers to join a community for collaborative problem-solving in AI character creation.
Takeaways
- ๐จ The process of creating a consistent AI-generated character involves three main parts: generating the character, refining the character sheet, and training the AI model.
- ๐ AI-generated characters often face the issue of non-reusability, but this can be mitigated by using character tournaments and control nets to generate variations of the same character in different poses.
- ๐ผ๏ธ Creating a character sheet with a variety of poses and expressions helps in maintaining the character's consistency and provides a solid foundation for AI training.
- ๐ก The use of control nets with duplicate open pose rings and stable diffusion can aid in generating the same character across multiple instances with varied poses.
- ๐ฎ For game development, a limited number of character poses may suffice, but for deeper character integration, a more detailed and dynamic character sheet is necessary.
- ๐ง The script outlines a method for creating a clean character sheet using Photoshop and reference images from Civic Ai, emphasizing the importance of high-quality images and varied poses.
- ๐ A spreadsheet is recommended to maintain variation in character poses while ensuring quality, which is crucial for the training process of the AI model.
- ๐ ๏ธ The script provides detailed steps for using stable diffusion and control nets to upscale and refine the character images, including the use of specific parameters and prompts.
- ๐จ The importance of cleanup and retouching in the image editing process is highlighted, as it helps in achieving consistency and clarity in the character's appearance.
- ๐ The creation of a data set from the refined character images is essential for training the AI model, with the script providing guidance on how to prepare and separate images for this purpose.
- ๐ The script concludes with the training of the AI model using KoiAS, detailing the renaming, captioning, and folder preparation steps required for effective model training.
Q & A
What is the main problem with AI-generated characters?
-The main problem with AI-generated characters is that they are non-reusable. Once you click generate again, the character is gone forever unless you use established names or trained models.
How can you create variations of a character in different poses?
-You can create variations of a character in different poses by using a character tournament for a turnaround of the same character or using control net with duplicate open pose rings to guide stable diffusion in generating the character in various poses.
What is the purpose of creating a character sheet?
-Creating a character sheet is the first step in the process of building a consistent character for AI generation. It helps in maintaining quality and variation in the character's poses and features.
What tools are used to create a clean character sheet?
-To create a clean character sheet, you can use open pose rigs in Photoshop, reference from Civic Ai, and extract poses directly from there. You can also use a preset of open pose rigs for the main turnaround.
How do you upscale a character image while maintaining quality?
-To upscale a character image, you can use high-res fix with a preferred upscaler like 4X Ultra sharp or lanzos, applying parameters such as 1024x1024 resolution, padding pixels of 42, and Sim fix activated for better results.
What is the role of regularization images in training a model?
-Regularization images provide additional flexibility to the model during training. They are variations of the character class in different poses, angles, expressions, and light setups, which help the model learn the character more effectively.
How do you prepare a data set for training a learner model?
-To prepare a data set, you start by cleaning up the upscaled image, separating it into individual poses, and then creating images from the cutouts with higher resolution. These images are then used to train the learner model.
What are some tips for captioning images during the training process?
-For captioning, you should use consistent naming and formatting for all images, describe them with relevant tags, and ensure the keyword that triggers your Laura model is placed at the beginning of each caption.
What is the significance of using a Discord server for character creation?
-A Discord server provides a community where individuals can share knowledge, experiment together, and find optimal solutions for character creation challenges, contributing to the collective understanding and advancement in the field.
What are some limitations of the current AI character creation process?
-Limitations include difficulty in achieving asymmetry, complex patterns, and creating non-humanoid or character sheets with items. The process also requires a significant amount of work and optimization to improve results.
How can you further improve a trained Laura model?
-You can improve a trained Laura model by retraining it using new generated images that show the desired improvements. This iterative process allows for continuous enhancement of the character's flexibility and precision.
Outlines
๐จ Character Creation with AI: An Introduction
The paragraph introduces a method for creating a consistent AI-generated character using a three-part process. It discusses the challenges of reusing AI-generated characters and suggests using character tournaments and control nets with duplicate open pose rings as solutions. The goal is to create a clean character sheet with variation in size, mix of headshots and full body shots, and dynamic and static poses. The paragraph emphasizes the importance of using a preset of open pose rigs and references from Civic Ai for the main turnaround, aiming to use this space efficiently.
๐๏ธ Refining the Character Sheet and Training Laura
This paragraph delves into the refinement of the character sheet and the training of Laura, an AI model. It describes the process of selecting the most consistent images, retouching for consistency, and preparing the character sheet for training. The paragraph also covers the use of photo editing software for retouching, the creation of variations using the extras option, and the importance of a clean character sheet with a white background for training. It concludes with the process of upscaling and creating regularization images for further training flexibility.
๐ Iterative Training and Dataset Preparation
The paragraph discusses the iterative process of training and dataset preparation. It explains the use of dynamic prompts and wildcards to create a list of concepts for the generation to choose from, aiming for simple images focusing on the main class. The paragraph details the process of creating a data set from the upscaled image, cleaning and separating it into different images for training. It also touches on the creation of images from cutouts and the optional step of adding environmental reflections and blurring for cohesiveness.
๐ Organizing and Tagging Training Images
This section outlines the organization and tagging of training images. It emphasizes the importance of renaming images to a consistent format and using tags for description. The paragraph explains the process of auto-captioning images and manually cleaning up the captions for better training results. It also discusses the creation of folders for training in Koias, selecting the appropriate model for training, and setting the training parameters. The goal is to prepare the data set and training configuration for effective character training.
๐ Training the AI Model and Observing Progress
The paragraph describes the actual training of the AI model using the prepared data set. It covers the process of setting up the training in Koias, selecting the right model, and adjusting training parameters for optimal results. The paragraph also discusses the use of sample images and the importance of monitoring the training progress to ensure the model is learning the character consistently. It highlights the potential for infinite improvement of the model by retraining with new generated images and encourages community collaboration for advancements in character creation.
๐ Evaluating Training Results and Future Improvements
The final paragraph evaluates the results of the AI model training and discusses potential improvements. It presents the outcomes of good training versus poor training, emphasizing the importance of a well-prepared data set and proper training techniques. The paragraph also talks about the limitations of current AI capabilities in character creation and suggests areas for future optimization. It concludes with an invitation to join a community for collaborative learning and improvement in AI character generation.
Mindmap
Keywords
๐กAI-generated characters
๐กCharacter embedding
๐กStable diffusion
๐กControl net
๐กCharacter sheet
๐กPhotoshop
๐กUpscaling
๐กData set
๐กTraining a learner model
๐กDiscord server
Highlights
The process of creating a consistent AI-generated character from scratch is divided into three parts, providing a step-by-step guide.
AI-generated characters often face the issue of non-reusability, where characters are lost once a new generation begins.
Using character tournaments and control nets with duplicate open pose rings can help generate the same character in various poses.
Creating a clean character sheet with variation in sizes, mix of headshots and full body shots, and dynamic and static poses is essential for efficient use of space.
Embeddings can be used to maintain character consistency, but the difference may not always be noticeable.
The use of a spreadsheet with a lot of variation while maintaining quality per pose is recommended for creating a character sheet.
Stable diffusion is utilized to generate the desired character, with an example of a woman wearing a kimono with pink hair and blue eyes.
Control net is used with a clean character sheet and a white background for better results in character generation.
The importance of using high image quality and experimenting with different prompts and models in stable diffusion is emphasized.
The use of CFG scale is suggested if the prompt feels ignored during the character generation process.
Freezing the seed and activating high-res fix helps to upscale the character sheet for better detail.
Creating variations of the character using the extras option in the control net makes the cleanup process easier.
Image-to-image upscaling is recommended if the white background is not achieved, using a reference with a white background.
The process of creating new poses from the base image and using them as references for further character generation is discussed.
Cleanup and retouching of the character images are necessary, depending on the importance of the character and the intended use.
The use of photo editing software for refining character images, such as adjusting colors, adding ornaments, and fixing inconsistencies, is described.
Upscaling the character images using specific parameters and software, like 4X Ultra sharp or lanzos, is recommended for better image quality.
Creating regularization images of the same character class in different poses, angles, and expressions provides flexibility for the final AI model.
The use of Dynamic prompts extension is suggested for creating a list of concepts and letting the generation choose one in every batch.
Preparing a data set from the upscaled and cleaned character images is crucial for training the AI model.
The process of training a learner model using KoiAS and the importance of proper data set preparation and model selection is discussed.
The potential of retraining the AI model using new generated images to improve character consistency and flexibility is introduced.
The creation of a Discord server for sharing knowledge and advancements in character creation and AI-related topics is announced.