I'll start with a small apology, I've decided to change how I generate the visual content for Workin' Out slightly. The original goal of this project was to deliver it quickly using a basic content development workflow in Stable Diffusion... however, as I've been working on a lot of images for chapter 3 I've been less and less satisfied with the character drift I've been seeing.

I define character drift as the the SD process deviating from the original character concept I described using the text prompts. SD and the base model have it's own architype that it's been delivering - whereas my vision of Gwendolyn is the image you see at the top of the project page.

That picture was one of the random edge case renders that SD sometimes delivers. As soon as I saw it come out of the pipeline I realized it was something special. Thereafter that image was that the true version of Gwen for the Workin' Out story. 

It's been very difficult to recreate that character through only text prompts... and the more tokens I add to the prompt, the further the image drifts away from the true Gwen. As an example, from the last post - this is the image I arrived at using the normal text prompt for Gwen, plus additional descriptive terms for her pose and mood and ControlNET for additional posing control:

I tried an experiment last night - I dusted off my old notes for creating a minimal LoRA for a character; something I had tried out a while ago. Starting with the 3 images I had of the true Gwen (a single pose with 3 expressions), I added a set of full body images with little facial detail and a few other pictures that had been generated along the way that matched pretty well with the character concept. After a few hours of training there was a character model that I can add to the rendering pipeline to strongly influence the look of Gwen. 

Using most of the text prompt for the image above, removing the character description and including the new character LoRA I was able to generate this image:

The dress style and posing aren't perfect, but the characterization is!

So with this new approach, I'm going back through all the Gwen graphical content and recreating them to be true to my vision of her. I'm afraid it's going to take me a bit more time than I'd planned to get all the content for milestone 2 ready... but I hope you'll forgive me. I think when I deliver it, the delay will have been worth it as it's less visibly jarring when the main character has a consistent look screen to screen.

I'll probably write a longer post about the set up and steps I took to build the LoRA, if anyone's interested. 

Meanwhile, I'll keep burning the midnight oil working to deliver Workin' Out!,

Cheers,

-Insomniac