A LoRA for Adriana Chechik.Process<ul><li>Images (71)<ul><li>Focus<ul><li>30 "full" body (waist/knees up)</li><li>17 upper body (chest and head)</li><li>18 close up (head and shoulders)</li><li>6 weird angles/poses (range from "full" body to upper body)</li></ul></li><li>Aspect ratio<ul><li>30 1:1</li><li>41 3:4</li></ul></li><li>Content (varied...)<ul><li>faces (1 eyes closed, half smiling, 1 eyeglasses)</li><li>lighting</li><li>clothing</li><li>makeup</li><li>background</li><li>pose</li></ul></li><li>Misc<ul><li>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.</li><li>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.</li><li>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.</li><li>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.</li></ul></li></ul></li><li>Captions<ul><li>All begin with "adriana chechik, a photo of a woman..."</li><li>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.</li><li>I do not describe things I do want to show up in the LoRA, like eye color, hair color, skin tone, body proportions, etc.</li><li>I have experimented with adding a fake word "ohwx" to the captions with varying results. I did not do so for this LoRA.</li></ul></li><li>Training Params<ul><li>model: DreamshaperXL</li><li>text_encoder_lr: 0.0004</li><li>unet_lr: 0.0004</li><li>learning_rate: 0.0004</li><li>network_dim: 256</li><li>network_alpha: 1</li><li>lr_scheduler: constant</li><li>optimizer_type: Adafactor</li><li>train_batch_size: 1</li><li>dataset repeats: 20</li><li>epochs: 10 (sometimes up to 12 if I have a highly varied dataset)</li><li>max_train_steps: 20 * 10 * # of images (so for this one, it was 20 * 10 * 71 = 14,200)</li></ul></li><li>How is it so small?<ul><li>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.</li></ul></li></ul>I'm sure I missed something here, so let me know if there's any other info that would be useful.

Files with SHA256: 9d74b5c09b97d3b670960d57a61dce13c733c568a1bc62bcd0ba55886da85497