This is my second LORA and I'm not super confident about it, but I wanted to post it here anyway to see if anyone can work some magic with it. This LORA is intended to in my experience, it can produce good results with really detailed prompts but can sometimes be pretty distorted. The dorm room scenes tend to work best, but it's possible to get a wide variety of environments depending on what checkpoint you use. It plays pretty well with other LORAs too.

Background on the dataset: This was trained with 16 images from a mixture of professional porn in a college setting (College Rules), as well as some amateur college party pics.

Suggested Checkpoint: This LORA was trained using BigLove as a point of reference. For best results, I would suggest using that model. If using with other checkpoints, you may need to add additional descriptors in your prompt. I tested with anime checkpoints and it didn't really work, so I would suggest using realistic models.

Strength: 0.6-0.8. From my experience, the LORA is pretty strong, and high weights can distort the results. I would try at 0.8 and lower it as you go. I didn't seem to get good results with going over 1 strength.

CFG: 5 worked fine for me. If I got a result close to what I wanted, I would sometime reuse the seed number and then adjust the CFG up or down to get a better result. I wouldn't go under 4 or over 8.

Sampler: DPM++ 2M SDE Karras worked well for me.

Steps: 25-30 produces decent quality at a good speed.

Prompt examples:

Prompt suggestions:

You can use the suggestions above or remix the prompts in my posted examples. Trying out different angles, such as "from above", "from below", "from side", as well as close-ups can produce different results. To generate full body, I suggest including something about feet, ankles, shoes, or socks into the prompt. This LORA can generate large crowds of people, but the more people there are, the more likely there will be distorted faces.

I would suggest resolutions 1024x1024 and higher. Portrait, landscape, and square can all produce different results. The higher the resolution, the more likely you will achieve clearer faces.