This is probably uninteresting for most people, just a fair warning. It's just me rambling.

I first started working on this gen in late September, and I quickly realized that I was honestly getting a bit out of my depth with this. The request for the Petronas Twin Towers in Malaysia was a bit complicated as the NoobAI checkpoint I usually use and am most comfortable with has no real knowledge of them, nor was there any SDXL-based Lora I could find for the Towers. At this point, I tried using a real image of the Petronas Towers to which I then realized I was going to need knowledge of the Towers in the model still because of inpainting any details that were going around or over the Towers, or I would lose some fine details that make the Towers what they are. I went looking further and I found an SD1.5 Lora but as you can imagine, SD1.5 is quite an outdated model and sure enough, I was just not happy with any of the results I was getting with it including even in inpainting. Funnily enough, this was my first time ever trying SD1.5 for myself since I started AI genning a bit later, but either way it was just bad. After that, I then tried to use a 3D model of the Towers for some reason, not that it changed my issue but hoping it would be stylistically "cheaper" to work on and maybe I could make something work with it with different details. It didn't really work out either, and I got as far as this before I realized I just didn't have the skill to mix the details and inpainting all that well and still get a full environment that looked good out of this. On top of that, I still had problems with any amount of inpainting on the Towers as the details of it just started being inconsistent when I tried.

I didn't quite give up but I did shelve this for a bit since I just didn't know what to do yet. At some point after that, I had eventually heard about Qwen Image 2.1 which actually came out just a few days before I started this request. I'm not too in the know when it comes to the technical side of AI models and I barely even keep up with any news about AI model stuff because I'm mostly only concerned with the furry side of it which doesn't have news too often, but my first thought was that it's a newer model with newer and possibly more capable data that maybe I should try it out for this. Turns out that was a good idea because it had pretty good knowledge of the Towers as you can see in this post. What also really came in handy was that Qwen Image 2.1 can generate transparent images. I ended up making good use of that feature because I needed to place Directo behind the Towers, so a transparent layer of just the Towers and surrounding buildings was perfect for achieving this and then I just did some more layer work to get the penis over the skybridge, as part of the request was having his penis resting over that skybridge. I generated a base character to assume Directo in my usual NoobAI model as that obviously has the furry knowledge I need for that part. I then generated a city background in Qwen as well to put behind Directo and so I then had 3 main layers of the image being character, background, and the Towers and accompanying buildings themselves.

I got quite far in doing this until I hit another major snag. For some reason, I wasn't able to inpaint with Qwen. It just kept crashing or erroring out and I just didn't know why. Idk if it was Qwen itself not being capable of inpainting, AMD's ROCm drivers being wonky (I have an AMD GPU), or Forge Neo not properly supporting it, but regardless, I couldn't use it. This was really disheartening because I was finally getting somewhere with this request and I obviously still couldn't use SDXL/NoobAI inpainting for anything on the Towers without it messing up something. This was my next dead end and I ended up pausing this gen again until I had any kind of bright idea or something that would help me out.

Fast forward a couple weeks later and then MOSS-Image comes out. Not only did MOSS also have decent knowledge of the Towers, but inpainting with it actually worked for me. Even better, MOSS is uncensored which actually made inpainting around the penis and Towers at the same time actually really doable. It's genuinely so lucky for me that two major models that would make this request possible for me not only came out around the time this request was made, but also came out so close to each other. Like, damn.

At that point, I had everything I needed. MOSS cleaned up most the areas surrounding the penis, skybridge, Directo's hands around the towers, and did the shadowing from the Towers on Directo with a blob I started. I'm surprised how close to the outline of the Tower that MOSS was able to do in that shadow, really. Inpainted some building details with MOSS, blobbed and inpainted Directo's OC details with NoobAI, and eventually I finally got to a point where I could say "I think this is done" and directo0265 gave me a thumbs up on it, so I'm happy with that.

This was by far the most demanding gen I have worked on yet, and it's nice to have a good challenge every once in a while, but man I needed some breaks from all the early frustration lol. I'm not a skilled artist and I would hardly consider myself an editor, but I feel like I at least managed to make something work. It's nice to see AI continually improving even for us local genners considering most of the industry focus has been on crazy huge models that basically nobody can run at home on consumer hardware. I genuinely couldn't have done this without Qwen 2.1 and MOSS and those just happened at the right time for me.