Sign In

Tag your training data for use in negative prompts!

Tag your training data for use in negative prompts!

I thought this was obvious, but apparently some people didn't know about it so I'll post an article on it. When training a LoRA, you should include not just tags applicable to the image you want the resource to be able to generate, but also tags that you want the LoRA to not generate so they can be used as a negative. The primary type of tag for this is artstyle, as the style of training images will have strong influence on generated images and many LoRAs will have a strong leaning (if not totality) toward a particular style, but there are a few other uses (obvious emotions/facial expressions). This is more important for some (3D) styles than others, but style tends to stick to anything.
Common style keywords (tell me in the comments if there's any more. I'm not sure of many for western media):
Anime: Anime screenshot, cel animation (if applicable)
Manga: Manga illustration, monochrome, greyscale (unless you've got a cover image or special page that's actually in color), comic (if things like panel divides are visible), Speech bubble, blank speech bubble (If applicable. Blank out any speech bubbles in an image editor if any are present. Text confuses SD)
Cartoon: Screencap, cel animation (if applicable), ???
CGI: 3D
Comic: Comic, speech bubble, blank speech bubble (if applicable) ???
Concept Art, promo images: Official Art (Note: This one is a very, very common positive prompt. Avoid using it if the image is low quality, even if it does actually qualify for this category)
Video game (sprite): Pixel art, simple background, white background (if on a clean background), game screenshot (if not)
Video game (model): 3d, game screenshot
Misc: Painting, illustration, photo (medium), traditional media
Multiple: retro artstyle, 1990s (style) (there's a tag for most decades)

This also helps if you have multiple styles in your TD, but want the user to be able to generate specific styles more closely. Pic unrelated.

6

Comments