Generating Unique Ethnic Characters with Stable Diffusion
The article discusses a technique called
Why it matters
This technique demonstrates how to leverage advanced AI models like SDXL to generate a diverse range of unique characters, which can be useful for various creative applications.
Key Points
- 1SDXL can generate a small number of repeating faces when prompted for a specific gender, age, and ethnicity
- 2The key is to use image-to-image generation instead of text-to-image, starting with a similar base image
- 3Iterate the process, using the generated image as the new starting point, and adjust parameters like denoise and cfg
Details
The article explains that to generate a quasi-infinite variety of unique faces, the author uses an image-to-image workflow with SDXL. Instead of relying solely on text prompts, the author starts with a base image that is somewhat similar to the desired character, and then iterates the process, using the generated image as the new starting point. This allows for more visual conditioning and results in more distinct faces. The author shares the specific parameters used, such as high denoise, low cfg, and the Karras scheduler. One downside is that the high denoise can sometimes lead to color splatters or stains on the face, which the author addresses by manually removing them.
No comments yet
Be the first to comment