[Preview] Fantasy Lady set and my workflow explained

I used Tifa's model to create the character, but the result doesn't quite resemble Tifa. Therefore, I've stopped referring to this character as Tifa and have decided to call her 'Fantasy' instead. This set was created using a brand new model I recently mixed. Generally, this model generates CGI-like characters with dramatic flesh and muscle shapes, something that a realistic style AI model cannot achieve.

I've been utilizing multiple processes to create ultra-high-definition images. My workflow is as follows:

1. I begin by generating images using generative AI, such as Stable Diffusion. This step usually yields a base image of 512*768 for portrait or 920*512 for widescreen. The purpose of this process is to test and determine which parameters can generate a good image. Each image typically takes about 20 seconds to generate. I generally go through the entire 360P pose library.

2. Once I've collected good parameters (usually from around 150 images), I use the generative AI again for image-to-image (img2img) upscaling regeneration. The AI reduces the noise in the base image slightly and uses it as a guide to regenerate the image at a higher quality. This process takes about 1-2 minutes per image. If the image is a portrait, an upscaling regeneration of 2-2.4 times is usually sufficient. The result is a 1024*1536 pixels image, which I refer to as an HD image.

3. When dealing with a widescreen image, the available space to draw the character might be smaller than in portrait mode, which can negatively affect the quality of the character. In such cases, you may not be able to see a clear face. Therefore, I have to redo the regeneration process, resulting in an image that is four times larger than the base image and has four times better quality. This process takes about 4-5 minutes per image, and the result is a 3680*2048 resolution image (four times the size of 920*512), which is very close to 2160P/4K. Due to memory limitations, I use a tiling script called SD Upscale to complete the process.

4. Each of the preceding processes may generate low-quality images, which I simply discard. By this stage, I typically have between 80-150 good images. To finalize them, I use another AI photo-editing program called Topaz Photo AI, a paid app, to upscale the HD image up to 6x. This is how I achieve an ultra-high-definition image of 10k or 11k.

In this specific project, my parameters for obtaining 4K widescreen images are:

• Base: 1200*672
• Img2img: 2x, 0.45 denoising
• SD Upscale: 1.8x, 0.2 denoising

Besides the generation time, I still have to spend 5-7 hours to check and tuning the images. I have already spend 17 hours on this set, and this is how my images are created

I haven't completed the final process yet, so please stay tuned for more updates.