HOME | DD

Description
I'm experimenting with using AI tools in a hybrid way, to pipe in original art and then apply style transfers (with minimal text entry/prompts), without dramatically overwriting the content & pose of the original image itself). There has been a marketplace for digital style-transfer tools (e.g., Photoshop actions and other tools) for a very looong time, so it's interesting to me to see if the AI tools can be scaled back to effectively do the same thing.

I purposefully started from an initially simplistic DAZ Studio render. Complicated scenes aren't really feasible for this type of image comparison. The starter orc image is a background-free portrait with lots of distinct facial detail and a high-angle / awkward head pose. The light sources are primarily from the back and the image-right (figure-left). In the examples I also included a couple of traditional (simple) filter-type actions, to compare / contrast them vs. what the AI tools do.

In ranking the results, I was trying to see how much "weight" was given to my input image. After all, I want to enhance the original with a style, not all-out replace scene elements. I ranked how closely the applications maintained the following elements:
1 - figure position
2 - lighting
3 - eye position
4 - scene elements

General results are:
- Midjourney really wants to take over the whole image and wreck all 4 evaluation criteria; the "image weight" param didn't have as much effect as I'd hoped; and Midjourney really didn't want to maintain the strange pose I gave in the source image (about 20% or less of the results maintained my original figure position)
- Wombo Dream (I recommend the browser interface, not the phone app) can preserve an original image fairly well for all 4 criteria and when given a limited prompt it (mostly) limits itself to applying the overall style to the source image;
- The *exact* eye position is a roll of the dice with all AI tools; however with a few variations I usually got a version that duplicated the original eye position.

Description
I'm experimenting with using AI tools in a hybrid way, to pipe in original art and then apply style transfers (with minimal text entry/prompts), without dramatically overwriting the content & pose of the original image itself). There has been a marketplace for digital style-transfer tools (e.g., Photoshop actions and other tools) for a very looong time, so it's interesting to me to see if the AI tools can be scaled back to effectively do the same thing.

I purposefully started from an initially simplistic DAZ Studio render. Complicated scenes aren't really feasible for this type of image comparison. The starter orc image is a background-free portrait with lots of distinct facial detail and a high-angle / awkward head pose. The light sources are primarily from the back and the image-right (figure-left). In the examples I also included a couple of traditional (simple) filter-type actions, to compare / contrast them vs. what the AI tools do.

In ranking the results, I was trying to see how much "weight" was given to my input image. After all, I want to enhance the original with a style, not all-out replace scene elements. I ranked how closely the applications maintained the following elements:
1 - figure position
2 - lighting
3 - eye position
4 - scene elements

General results are:
- Midjourney really wants to take over the whole image and wreck all 4 evaluation criteria; the "image weight" param didn't have as much effect as I'd hoped; and Midjourney really didn't want to maintain the strange pose I gave in the source image (about 20% or less of the results maintained my original figure position)
- Wombo Dream (I recommend the browser interface, not the phone app) can preserve an original image fairly well for all 4 criteria and when given a limited prompt it (mostly) limits itself to applying the overall style to the source image;
- The *exact* eye position is a roll of the dice with all AI tools; however with a few variations I usually got a version that duplicated the original eye position.