Join my new Discord here! https://discord.gg/cnVvnkgtPb

Belle Delphine (Face & Body LoRA) - the infamous e-girl. This LoRA includes Belle Delphine's face AND body. Use the trigger words to get the desired result. Infinite Belle pics await you.

NSFW Previews: https://jpg.pet/aihotgirls

Read Me!

TWO VERSIONS:

• Belle Delphine RHM v1: Trained on Real Hot Mix, my own mix. I use this LoRA for extremely high quality photorealistic results

  • For txt2img & img2img: LoRA weight between 0.9-1 for best results

• Belle Delphine SD1.5 v1: Trained on SD 1.5 base model, compatible with all models

  • For txt2img & img2img: LoRA weight between 0.5-1 depending on checkpoint model for best results (lower for anime/non-photorealistic models)

Recommended Settings

• Hires fix: 0.2-0.3 for face, 0.2-0.4 for full body

• Face is best at 1:1 resolution (make sure inpainting is 1:1 resolution)

  • Other resolutions work as well (see Training Info: resolution sizing for training dataset aspect ratio info)

• 'makeup' is a trigger word. Put it in negative and use 'brown hair' to get Belle with no makeup

• This LoRA is NSFW skewed. Use 'clothes' in positive and '(nude:1.5), (breasts:1.5)' in negative if you want SFW pics

• SOMETIMES:

  • If generating on not 1:1 aspect, her face may become slightly elongated. Maybe 1 in 5 images

  • 'pink hair' occasionally makes humongous hair. Not sure why this happens. I use 'huge hair, long hair' in negative if that seems to be happening with my prompt

Training Info

• 5 epochs, 25 repeats, 36 images

  • 21 brown hair, 10 pink hair, 5 blonde hair pics (if unspecified, brown hair will usually default)

  • 22 face, 14 full body pics

• For resolution sizing, enable buckets = true (25 = 1 image)

  • bucket 0: resolution (384, 640), count: 25

    bucket 1: resolution (384, 704), count: 75

    bucket 2: resolution (448, 576), count: 100

    bucket 3: resolution (448, 640), count: 25

    bucket 4: resolution (448, 704), count: 100

    bucket 5: resolution (512, 576), count: 25

    bucket 6: resolution (512, 640), count: 75

    bucket 7: resolution (576, 576), count: 450

    bucket 8: resolution (640, 448), count: 25

Background

I've been experimenting with LoRA training and seem to have found a sweet spot for epic photoreal accuracy for people. I didn't realize there were already so many Belle LoRAs while making this, but wanted to share anyway. I believe this one is the BEST one on Civit.