⚖️ Weight:    1

🔑 Trigger word:    caroline

📌 Key words

caroline, 1girl, star eye, green eyes, blue jacket, white shirt, waist wrap, black shorts, armband, ahoge, pink hair, multicolored thighhighs, white knee boots

🎲 Optional prompts

gold center stripe, pankou knots, silver_buckle

⚠️ Important Note

✦ A lot of character details are well-trained in this LoRA. Key words have good performance mostly. Add optional prompts only when details fail to generate correctly.

✦ Note that full-body images at 1024 resolution cannot achieve perfect details due to checkpoint limitations. Please enable Hi-res Fix first to get a usable base image. Then you can use inpainting to refine details that are not good enough. Keep the LoRA loaded during inpainting, and only input the trigger word for inpainting. Add targeted optional prompts only for elements that are difficult to generate.

✦ This LoRA is specially trained for the tiny star highlights in the character's eyes. To avoid conflicts with the common star in eye expression tag, do not use relevant Danbooru labels. You must use these training tags: green eyes, star eye. Add star pupils to your negative prompts, and adjust the weight of star eye as needed. Star highlights cannot be rendered at low resolution; inpainting is required for non-close-up shots. If you want to create a classic starry-eye expression, remove star eye and use standard Danbooru tags such as star in eye, 🤩.

✦ It may be necessary to lower the LoRA weight when generating alternate outfits.

✦ The quality of inpainting depends heavily on personal skills, including mask, denoising strength, plugins and other parameter settings. Beginners may need more trial.

💡 Common Tips

✦ Using Hi-res Fix can achieve better detail performance.

✦ Load the LoRA and inpaint with the trigger word for perfect results.

✦ If the result is very different from the intended character, the LoRA may not have loaded properly or its weight is incorrect.

✦ If the detail quality is unsuitable, try increasing the weight (without overfitting), add more precise descriptions as needed, or include negative prompts for elements you want to remove.

✦ Different base models may render details differently, and inherent model biases are difficult to avoid. Adjust your prompts if the generations are not ideal.

✦ All LoRAs trained on the Illustrious base model are theoretically compatible with NoobAI.