The basic idea of this LoRA is to create cars with widebody.
When used with car LoRAs or checkpoints that could draw cars, it should be able to add some kind of widebody kits onto the cars.
All cars are categorized into 6 different types:
Sports Car
Super Car
Sedan
Hatchback
SUV
Truck
Recommended Weight: 0.6~1.0
Prompt Format: color (optional) + widebody + body type (see above)
Sample prompts to start with:
((masterpiece,best quality)), ultra-detailed, blue widebody sports car, vehicle focus,
Description
FAQ
Comments (7)
hey man, how did you categorize as different car segments? is this done during captioning? if so, what kind of captioning?
Yes, it is done during captioning (if what you mean is the process where you add tags to each training data (pic)).
As for the categorization, I simply put car pics under the same segment into the same folder. But you do need to determine which pic belongs to which segment, as the tagger itself will only recognize thoes pics as "cars".
@fallenL ah i see thank you for your answer. just a quick follow up, what do you mean by put car pics under the same segment into the same folder ? Do you put all SUV, coupe, sedan pics in same folder, but then use the tag to put "SUV" "coupe" etc during captioning?
@sjyoh93481 Yes and no.
Let's say you are training a car concept LORA, and you have two different concepts.
To put them into one single LORA, you need to create two sub folders under your lora's folder (one for concept A and one for concept B). Then for every pic that is categorized as concept A, you put it into concept A's folder; vise versa. This process could be done at anytime before the training.
Basically this is like putting books (pics) back to the sections (concept folder) that they belong to in a library (LORA).
@fallenL Amazing. learned something new thanks !
Does the sports car in your model only appear in the training set? I am trying to use the training set of sports cars of different shapes to train lora. I hope that when the prompt word input sports car, sports cars of different shapes will appear, but there will be structural errors.
Sorry I don't understand what you mean for your question?
As for the issue that you mentioned, I would suggest using different prompt word(s) for your Lora concept, as most of the checkpoints could recognize the concept of "sports car" without any additional Loras. (What I'm trying to say here is try to avoid naming your prompt(s) with word(s) that could often be found in checkpoints, unless the training dataset is large enough to outweigh the same prompt(s) in a checkpoint.)