【Stable Diffusion】背景を上手く生成させるためのコツと呪文集をまとめて紹介
TLDRThe video provides a comprehensive guide on generating high-quality backgrounds using AI models, specifically 'anylora' and 'anythingV5' from the Civitai distribution site. It emphasizes the use of specific prompts and negative prompts to create background-only images, such as adding 'background only:2' and excluding terms like 'character' or 'person'. The importance of image size for quality is highlighted, with suggestions to use 'hires.fix' or set dimensions to 1024 or higher. The video also discusses the impact of different sampling methods and the need to experiment with various models to achieve desired styles. It mentions that certain types of backgrounds, like natural landscapes, are easier to generate than specific indoor settings. The guide offers various 'spells' or keywords for generating different types of backgrounds, such as 'morning', 'sea', 'japanese-style room', and 'sun', along with tips for adjusting image size to improve generation outcomes. It concludes by encouraging viewers to subscribe for more informative content on AI-generated art.
Takeaways
- 🎨 **Background Creation Focus**: You can generate only the background for use in illustration collections or game development materials.
- 🔍 **Model Selection**: Different models can specialize in generating backgrounds; 'anylora' and 'anythingV5' from Civitai are suggested.
- ❌ **Background-Only Prompts**: To avoid generating characters, use prompts like 'background only:2' and negative prompts such as 'no character'.
- 🙅♂️ **Omitting People**: Suggests using 'character, person, human, people' as negative prompts to exclude people from the generated images.
- 🖼️ **High-Quality Backgrounds**: Use phrases like 'masterpiece, best quality, detailed background' to enhance the quality of the generated background.
- 📏 **Image Size Matters**: Increasing the image size to 1024 or more on one side, or using 'hires.fix', can improve the quality of the background.
- 🔄 **Sampling Methods**: Experiment with different sampling methods as they can affect the quality and style of the generated background.
- 🏞️ **Natural vs. Indoor**: Natural backgrounds are easier to generate, while indoor settings like buildings can be more challenging.
- 🔄 **Model Training**: The ability to generate certain backgrounds depends on the model's training; using multiple models can increase success.
- 🔢 **Image Size Adjustments**: If the desired background isn't generated, changing the image size can sometimes achieve the desired result.
- 🌅 **Time and Weather Spells**: Specific background spells like 'morning', 'day', 'evening', and 'sun' can help generate backgrounds with the desired time of day or weather.
Q & A
What is the main purpose of the video?
-The main purpose of the video is to introduce a collection of tricks and background spells for successfully generating backgrounds using 'anylora' and 'anythingV5' from the Civitai model distribution site.
Why might someone want to generate only the background in an image?
-Some people might want to generate only the background to create an illustration collection or to use it as material for game development.
How can you increase the variation of scenery in the generated background?
-You can increase the variation of scenery by using different models, changing the image size, and experimenting with various background spells and sampling methods.
What is a recommended negative prompt to omit people from the generated background?
-A recommended negative prompt to omit people is to use terms like 'character, person, human, people'.
How can you ensure the background is of high quality?
-To ensure high-quality backgrounds, use the spell 'masterpiece, best quality, detailed background' and consider increasing the image size to 1024 or more on one side or use 'hires.fix'.
What should you do if the generated background doesn't match the desired style?
-If the background doesn't match the desired style, you can try changing the sampling method or the model used for generation.
Why is it recommended to use multiple models when generating backgrounds?
-Using multiple models is recommended because it increases the likelihood of generating a wide variety of backgrounds, as different models have been trained on different datasets and may produce different results.
How can changing the image size affect the generation of the background?
-Changing the image size can affect the generation of the background by potentially making it easier to generate according to the spell of the background, especially if the size is set to 600 vertical or 600 horizontal.
What is a specific example of a background spell for an early morning setting?
-A specific example of a background spell for an early morning setting is to add 'morning' to the prompt, and for more specificity, combine it with another spell like 'morning, sea'.
What is the advice for generating indoor scenes like a Japanese-style room?
-To generate an image of a Japanese-style room, enter 'japanese-style room' as a prompt, and consider using a Japanese-style human model or background model to enhance the quality of the generated image.
What is the final advice given in the video for generating backgrounds that are not coming out as expected?
-The final advice is to try changing the image size or experimenting with different models, as these adjustments can sometimes lead to successfully generating the desired background.
Outlines
🎨 Background Creation Techniques
This paragraph discusses the desire to generate only the background for various purposes, such as illustration collections or game development. It introduces the use of 'anylora' and 'anythingV5' from the Civitai model distribution site to achieve this. The text emphasizes the use of specific prompts and negative prompts to generate backgrounds without characters, suggesting the use of 'background only:2' and excluding terms like 'character, person, human, people'. It also advises on enhancing image quality by using phrases like 'masterpiece, best quality, detailed background' and adjusting image size or sampling methods. The paragraph acknowledges that not all backgrounds can be generated with a single model and suggests using multiple models or lora models to increase the chances of success. It concludes with tips on generating specific types of backgrounds, such as early morning scenes or indoor settings like a Japanese-style room.
🌄 Outdoor Background Generation
The second paragraph begins with a note on the potential blurriness of generated images but expresses optimism in creating beautiful images using Japanese-style models. It then transitions into discussing outdoor background spells, emphasizing the limitations of a single model to generate all types of backgrounds. The paragraph suggests using multiple models or a specialized model for generating backgrounds if the desired image is not produced. It also mentions that changing the image size can sometimes lead to successful generation. The video concludes with a summary of the techniques and spells for creating good backgrounds and encourages viewers to subscribe for more information on generated AI.
Mindmap
Keywords
💡Background Generation
💡Anilora and AnythingV5
💡Background-only Prompts
💡Negative Prompts
💡Image Quality
💡Sampling Method
💡Model Training
💡Checkpoint Models
💡Image Size
💡Background Spells
💡Indoor and Outdoor Spells
Highlights
You can generate only the background using certain models.
Adding 'background only:2' to the prompt helps generate backgrounds with fewer people.
Using negative prompts like 'character, person, human, people' can omit people from the generated images.
Different models may require changing prompts or trying different models if you always get people in the generated images.
To improve background quality, use phrases like 'masterpiece, best quality, detailed background'.
Larger image sizes, such as 1024 or more on one side, can enhance the quality of the generated background.
Using 'hires.fix' can double the image size, further improving quality.
The ideal style of the background may depend on the sampling method used.
Experimenting with different sampling methods can help find the best one for a particular model.
It's not possible to generate all types of backgrounds with a single model.
Natural backgrounds are easier to generate, while certain indoor settings like buildings can be more challenging.
The ability to generate a background depends on the model's training.
Including multiple models or using a model that specializes in generating backgrounds can improve results.
If the desired background isn't generated well, adjusting the image size may help.
For early morning backgrounds, add the 'morning' spell to the prompt.
Specifying weather conditions, like 'morning, sea', can make it easier to generate the desired background.
Setting the image size to 600 vertical or horizontal can help generate backgrounds according to the spell.
Adding a 'day' spell can generate daytime images, while an 'evening' spell is used for sunsets.
Combining 'sun, sea' with the background spell can make it easier to generate images with the sun.
For indoor settings like a Japanese-style room, entering 'japanese-style room' can help generate the desired image.
Outdoor settings can also be generated using specific spells tailored to the desired environment.
In summary, while it's possible to create only the background with a model, using multiple models or a specialized one may be necessary for certain backgrounds.