Adriana Chechik SDXL

A LoRA for Adriana Chechik.

Process

• Images (71)

  • Focus

    • 30 "full" body (waist/knees up)

    • 17 upper body (chest and head)

    • 18 close up (head and shoulders)

    • 6 weird angles/poses (range from "full" body to upper body)

  • Aspect ratio

    • 30 1:1

    • 41 3:4

  • Content (varied...)

    • faces (1 eyes closed, half smiling, 1 eyeglasses)

    • lighting

    • clothing

    • makeup

    • background

    • pose

  • Misc

    • I try to exclude any images that have a busy/complex scene/background. Abnormal clothing, hand gestures, etc. are cropped out when possible. My rule of thumb is that if I wouldn't want the image to be generated by the LoRA, I don't include it in the dataset. There are some exceptions to this rule, but it is a good starting point to trim the dataset.

    • As many duplicate clothing items, facial expressions, poses, pieces of jewelry, etc. are excluded as possible, but it can often be hard to avoid this.

    • Images are cropped by hand and left at whatever # of pixels achieves the desired final image. They are kept to 3:4, 4:3, or 1:1 aspect ratios.

    • Many others have commented that 71 images is unnecessary, and that 20 or so will do. I prefer to be in the 40-80 range.

• Captions

  • All begin with "adriana chechik, a photo of a woman..."

  • I describe the clothing, jewelry, lighting, pose, angle, background, facial expression, makeup, and any other information I do not want showing up in the LoRA gens (abnormal hair color, for example) in sentence form.

  • I do not describe things I do want to show up in the LoRA, like eye color, hair color, skin tone, body proportions, etc.

  • I have experimented with adding a fake word "ohwx" to the captions with varying results. I did not do so for this LoRA.

• Training Params

  • model: DreamshaperXL

  • text_encoder_lr: 0.0004

  • unet_lr: 0.0004

  • learning_rate: 0.0004

  • network_dim: 256

  • network_alpha: 1

  • lr_scheduler: constant

  • optimizer_type: Adafactor

  • train_batch_size: 1

  • dataset repeats: 20

  • epochs: 10 (sometimes up to 12 if I have a highly varied dataset)

  • max_train_steps: 20 * 10 * # of images (so for this one, it was 20 * 10 * 71 = 14,200)

• How is it so small?

  • After training is complete, I am left with a 1.7gb safetensors file. I use the kohya gui to resize the lora with a rank of 256. This spits out a ~18mb safetensors file that is nearly identical to the 1.7gb file in practice.

I'm sure I missed something here, so let me know if there's any other info that would be useful.