santa hat
deerdeer nosedeer glow
Sign In

I'm thinking about training a LORA, what kind of images I should avoid using?

Refer to title, I do want to try training a LORA, but I'm not sure what images I should avoid using. Like... I should I avoid the lineart, flat colors, can I use official shots from their show, pretty much a general guide. Mostly because of the character I'm doing lacks any detailed art.

I also think that this will help with new people trying to pick out images for their own LORAs.

4 Answers

Hey there! It kinda of depends on what exactly you're trying to train.

Say if I was personally going to be training a TV show character, I would do my best to get as many high-quality screenshots as possible. Make sure the images aren't blurry and there aren't any watermarks or anything weird in them that the AI will pick up. Next, if this character has a lot of drawn fan art, I would then select some high-quality drawn fan art. Again, remove any watermarks or anything weird that you do not want the AI to pick up...then train on those initial images. For a character, I would aim for around 20-40 images, at least that's what I have had success with personally.

Sometimes, after training the character likeness or style, whatever I'm trying to do, doesn't come out consistently. What I will then do is generate as many images as I can that come out relatively consistent...and then I will do another train with them added to the data set. justTNP has a pretty good guide over gathering character data sets for LoRAs here too. :)

Simply put

Get clear, highres image of the same concept

Avoid blurry ,lowres multi concept image one folder

For Official AnimeTV Like i recommend to go on, you have Anime/Film/Series, They have generally a good Quality output for Screenshots. (1080P).

For High Quality Original Art, & Are the Two Best for me Personally, followed by Gelbooru, tho Booru site are good, you need to filter somes Artwork with Tags and removes one you don't like by Quality & Prescision. (Depending on what you want).

Avoid multiple character's in the same images, tho it's not a problem if they are in Minority, But generally go for where the Character's is Purely Solo.

Avoid also Blurry images, Pastel/Paint Artwork, If you can try to go with the most Sharp Artwork. (Because AI Sometimes don't understand some concept when Artwork is Blurry/Painted).

Minimum of images to gather, I will say probably 20-30 is the Minimum, for a good datasets 50+ & If you can go higher it's better. (Since it's add more Flexibility/Poses/Angles).

Always Priorize Quality over Quantity, this is the most important in a Datasets, because that will define a good part in you're Lora's. (Quality/Sharpness). I prefer 20-30+ Artwork of the Best Quality possible than 100+ Artwork with Medium Quality.

It's not all but, This is a very good Starting point. GL 👍.

You already have a lot of information for what kind of images you shouldn't use in general.

My personal recommendation for character LoRA is to use everything with a good resemblance, irrespective of the artstyle. If there are chibis, monochrome or other more unique representations it only makes the LoRA more robust.

For one of my character LoRA I used a pretty big set of (animated) movie screenshots. It appears to be quite flexible but takes a big hit whenever it is combined with more cartoony styles.

Somewhere in the range of 20-30 images should be enough if they are high quality. Each, if you also want to train multiple costumes. But if you can get it, aim for at least 50.

(A tip for stylistic LoRA/LoCon - if you recreate someones style you should be aware that those tend to evolve. Don't just grab everything with high quality but differentiate. )

Your answer