pixai.art/artwork/1753506527005787943

Merging to DIY LoRAs is interesting like mixing drinks; it can taste great and sometimes disgusting.

It does give unique looks but is rough to control. For one, their eyes often come out badly. The colors are often off of what the model supposes to be as well. The model I use for this art definitely is not like this but it looks pretty cool nonetheless. A lot of negative fixes is needed during the process for merged LoRAs, it seems.

It starts from here:

And here is the finished-ish product:



Actually it is more complicate than I thought it would. Inpainting (drawing on a certain parts of the image) is tricky. AI will do the jobs but it may not be what you want especially bigger inpainting instead of small one to fix few details. Using outside software is needed too.

I used an image editor (Krita) make a space for inpainting, painted the background first, and then draw another girl using inpainting on top of the empty space. The new girl image was darker the original so I have to re-adjust many times to bring it up. The one I ended up publishing is not the best too.



pixai.art/artwork/1753193818648844438

But I stick to it because the "better" one is very messy after being re-painted many times and it looks quite different than the first, one-girl version.

Trying to install Stable Diffusion on my PC. It has been a huge hassle thanks to my AMD GPU. SD is built for nVidia chips as they are engineered for AI in general. AMD has been trailing for a long time in this category. There are ways to work around it but still.....

I spent last 2 days learning how to brew pictures of multiple different objects in one image with only prompts. I find out that without other tools it is next to impossible to do it and get exactly what you want.

First, I managed to create an excellent base image of 6 girls with different colored dresses. The purple one on the far right was a bit tough to get the color right since it blends with the background (should have based it with white color instead of blue) and turns out white.



Then I touched up and upscaled it. It comes out pretty nicely if you look far enough with a lot of minor errors all over the place.



On another route off the base image, I used After Detailer (adetailer) extension to touch up different parts of the image many times. Also, I did some inpainting to correct major errors (hands, legs, etc) and ran it through adetailer again. It came out quite beautiful.



After using Ultimate SD Upscaler, it looks quite perfect.



I want to compare working on A1111 locally to working on pixai.art, my first experience with AI art, as well. So I uploaded the base image onto pixai.art and upscaled it.



They don't look that different.

I also accidentally did it with a different model, but it looks quite alright.



Lastly, I did the same with the touched-up image. The result is....



Better, as there must be more enhancers used for its ultimate enhancer function. Still, the ultimate enhancer function there though can produce terrible, error-prone results as well.

In conclusion, drawing multiple characters onto the same image in AI is quite different. Also, A1111 and pixai.art both have pros and cons. I can micromanage on A1111 better but pixai.art is much more convenient and easier to get highest quality results.

I haven't done it on pixai.art much lately as I have been enjoying with Automatic1111 and ComfyUI on my local machine. But here is one of nice ones on there.

I use a part of lyrics of The Chemical Brothers' "Galvanize" as prompts for an AI art. It comes out amazingly.



Damn, the details of multiple humans after upscaling are very nice!