309 Images: Step-by-Step Generation Walkthrough [NSFW]

309 Images: Step-by-Step Generation Walkthrough

Check out my original post and article detailing my creative process. This article is a long form response to some of the questions people left in comments.

TLDR: An article detailing some of my usual creative practices and a detailed walkthrough of a recent post and the obscene number of images I generated before I landed on the final batch of images I posted.

My General Rules

Some of this might be things I've posted in the past, so apologies if you've already heard me on this soapbox. All images were generated in Civitai and while I will often use OpenArt for post production edits, none of the images for this project were taken outside of Civitai.

Quantity first, then quality

My general goal is to make at least one post with 20 images a day. Repetition is one sure fire way to get good at any skill or ability be it sports, programing, storytelling. We can often get so anxious putting out content and wanting it to be perfect that we become paralyzed. There comes a point in the creative process that you're getting diminishing returns in that you could easily spend 2 extra hours to improve the quality of you content by 5% at most. So I will usually give myself a day to come up with an idea for a post, I generate the images, do what touchups I can, and then schedule the post and get ready for the next day. The benefit of this process is I seek out new Checkpoints and LoRAs and I put them together to see what pops out. It can be time a time and buzz intensive process, but I enjoy trying to come up with novel images that people haven't seen before.

Play with Newer Checkpoints & LoRAs

Before I started publishing my Hundreds of Lulus posts I had the tendency to rely heavily on just a handful of Checkpoints for more of my generations. It wasn't until I started making it a point to test out newer models when they posted that I realized that it was a great way to support other creators on the site while also helping me discover interesting styles. I'll often also look for the newest style LoRAs or look for LoRAs that perhaps are cool and underused. This is also a bit of a time and buzz suck because a combination of LoRAs couple with one Checkpoint makes trash while pairing it with a different Checkpoint churns out some great images. In the end it's all about experimenting and trusting your gut if you think something looks not quite right or really awesome. This is also the point where I'll start tweaking the intensity of the models and trying to figure out the advanced settings.

KISS Prompts to Start

In the beginning I will usually start with a very basic prompt (Keep It Simple Stupid). Since the first objective is finding the right Checkpoint and LoRAs I'm more interested in getting a sense of the raw nature output before I start positioning the scene. Some models end up always placing the characters in the exact same position, or dressing them in the exact same style of clothes, or are to weak or overpowering when combined with other models. Once I get a good feeling for the Checkpoint and LoRA I then start getting more specific with the prompt, I start adding more details and look for the right prompts and negative prompts I need to add, remove, or adjust in order to finally get the effect I'm going for. 

Decide on Variety of Images or Variety of Checkpoints

With everything dialed in I will usually choose to produce all 20 images using a single Checkpoint but adjusting the prompts slightly in order to get image variety. Or I will take the same prompt and run it through multiple Checkpoints to see how flexible the concept is. Sometimes I'll do something in the middle where I select 2-4 Checkpoints and adjust the prompts slightly because sometimes you strike gold and end up with a fluke of an image that makes the whole process work it.

Work to 20

Finally, I will usually end up with 5-10 images I really like but I push myself to create 20. It goes back to the concept of repetition. There came a point when I started generating AI images where I was basically copying and pasting the recipes of other creators and perhaps changing the hair color or the background setting. There is absolutely nothing wrong with that, but I knew that the only way I was going to learn and get better is if I tweaked and adjusted until I got to the point I was creating my own recipes other people would want to use. The reason creating 20 images is key for me is it forces me to push and tweak everything I've just built. It gives me a sense of accomplishment that I created something that's not brittle but that can bend a little.

Start to Finish

Below contains a thumbnail for all the images I created for this project. To be clear, this was a really long and difficult project with me testing out a lot of combinations until I settled on something I liked. The final images are mostly the HiRes Fix images I used for the original post. In the end it came to around 310 images and ~1.5K blue buzz. Since I have a membership I can use my amassed blue buzz for the NSFW images, otherwise I would have likely stoped this project much earlier. It was an obscene amount and I wouldn't want anyone thinking this is my normal everyday process. I've also created this post if you're interested in diving deeper into the assets and prompts I used.

My Initial Concept and Flow

Earlier I had posted an article where I clearly said I was going to keep making smut, so I figured I needed to create something smutty, but visually interesting. I had not done anything smutty with Lulu lately so I started off by testing her with some of the newest Checkpoints.

From there I found a LoRA that created psychedelic renderings of characters, but all the images looked too cookie cutter.

I switched to a different psychedelic LoRA and was a bit more pleased with the output.

I added Lulu back in to the generations.

I then ran through a large number of Checkpoints and while the images were interesting, I also found them boring.

I then went looking to amp up the psychedelic imagery again.

I then figured I'd add on a sketch effect.

The sketch effect then transitioned to a ballpoint pen effect.

Eventually, I tested a line paper effect. However, it just wasn't working like I had hoped.

I ditched the line paper and had the idea of adding an ink wash effect and that's when I finally felt like I was getting close.

I swapped out the blue ballpoint pen ink for black and started adjusting the colors of the wash. I really started to like the images that showed Lulu emerging from a pool of ink.

I tried to remove the skin tone so that ideally the image would be monotone other than her pink hair and hazel eyes.

I liked how some of the images showed her clothes with ink running down her thighs and covering her hands. I removed the shoes and adjusted the prompts to remove the clothes and replace them with patches of ink. I also swapped out 'kneeling' for 'pinup pose' and added 'pool of ink' while also switching between landscape and portrait.

Getting happy with the results I kept tweaking the prompt to make sure that Lulu had ink on her legs and feet.

I finally felt happy with the prompt and started testing the prompt out across multiple checkpoints. I'd generate two portrait images and two landscape images per checkpoint.

Once I had my first batch of images I'd pick the best portrait and landscape image for each checkpoint and do a HiRes fix to get the final image for my post.

I then pulled the final 20 images together and shared the end results with you fine people.