- LoRAs are smaller files that you use in combination with existing Stable Diffusion checkpoints (models designed for generating images of a general/specific genre), that introduce a new subject to the image. This can be anything from a character to an art style to a pose.
- When using LoRAs, too much emphasis on a lot of other keywords are unnecessary and take away focus from the rest of the keywords related to the lora
- Below is an example of two similar prompts. The LoRA-focused one has less brackets.
Keyword-Focused
lora:AllFoursFromAboveV1:1, all fours, top-down_bottom-up, from above, looking up (((manga line work art style:1.4))), ((full body:1.4)), (highres), sfw (cute goth girl:1.4), (white hair:1.4), ((full body), tattoos, (choker, collar), best quality, hyper detailed, masterpiece, detailed face, orange eyes, highest details, orange halo, orange backlighting, light rays, (high contrast), (colorful)
LoRA-Focused lora:AllFoursFromAboveV1:1, all fours, top-down_bottom-up, from above, looking up, manga line work art style, full body, highres, sfw, cute, goth girl, white hair, tattoos, choker, collar, best quality, masterpiece, detailed face, orange eyes, orange halo, orange backlighting, light rays, high contrast, colorful
- Adjust emphasis with a single bracket one by one as needed.
- Things to note:
- Deliberate models make great quality outputs, but not necessarily the right pose, ReAL models attempt the pose better but the anatomy is off.
- LoRA works good with ReAL models, but seems to do better with the Deliberate model.