SDXL - Asphalt Potholes

Took a few pictures of my street, which is in perfect condition as you may assume, mixed it with a few AI generated images and voila. Made me look like an idiot though, because i had to explain to the neighbors that i don't take those pictures to complain at the city hall about potholes.

The size of the LoRA is another thing i wanted to test. I never used the resize feature in Kohya and i downloaded a LoRA that was way to big ( like 800MB ), so i tried to resize it. I thought setting it network rank 32 would lead to a file that is 200MB in size, but the file was just like 25MB. Was a bit skeptical about it, but after testing the same prompt with the same seed there was almost no difference. Given my ever growing collection of checkpoints and LoRAs, that could help to save some space. Unfortunately i can't do it to LoRAs that are made by someone else, cause it changes the Hash-ID and cross-posts wouldn't happen ( which is already kinda hit or miss anyway )

I know i could train a LoRA from the get go to be smaller, but some settings don't like a rank that is to low, which makes the LoRA unusable or it won't train at all.

I knew about the file-size stuff for quite some time and that it sometimes doesn't matter how big the file is. Even textual inversions are like only 5-25kb and have a huge impact, but i guess that depends on how big and complex your dataset is.

• Even without this LoRA using the prompts in some models will get you the same result, but i was focusing more on the depth of the potholes from different perspectives and the deep black color of fresh aslphalt against old weathered asphalt.

• Overtrained. I had quite a few pictures and the model picked the concept up right away. Usually i would let the trainer spit out a model every few hundred steps, depending on the size of the dataset, but i was testing a more complex model before i started this one and it was set to overwrite the last file every time a new file is made, so i lost the ones with less steps. What that means is that it will overpower the prompt - Low strength or long/descriptive prompts are recommended.

• The moss stuff is only because i used it cause it looks nice in my opinion, wasn't trained on that, but some pictures had weeds in them, so i guess the AI does the rest ( green tings in a picture could mean anything unless you describe what it is in the capitations )