LoRA for subtracting or adding color to lines.
This is more of an as-is model that I created with two images in 60 steps on ΣΙΗ 1.5 so I am not promising that it is well made.
Process
Made using: TrainTrain
Dataset: Left as Origin and right as Target
The differences are subtle because even small isolated differences have a large impact and this method is less destructive when resizing.
TrainTrain .json settings: Truncated based on what I changed from default.
{ "network_type": "lierla", "network_rank": "4", "network_alpha": "4", "train_iterations": "60", "train_batch_size": "1", "train_learning_rate": "0.0001", "train_optimizer": "AdamW", "train_lr_scheduler": "cosine_annealing", "train_lr_scheduler_settings": "eta_min=0.00002", "use_gradient_checkpointing": true, "network_blocks(BASE = TextEncoder)": [ "IN00", "IN01", "IN02", "IN03", "IN04", "IN05", "IN06", "IN07", "IN08", "IN09", "IN10", "IN11", "M00", "OUT00", "OUT02", "OUT03", "OUT04", "OUT06", "OUT07", "OUT08", "OUT09", "OUT10", "OUT11", "OUT05", "OUT01" ], "network_strength": "1.0", "network_conv_rank": "0", "network_conv_alpha": "0", "network_element": "Full", "network_train_text_encoder": false, "image_max_ratio": "2.0", "train_min_timesteps": "400", "train_max_timesteps": "800", "train_loss_function": "Smooth-L1", "train_lr_scheduler_num_cycles": "1", "train_lr_scheduler_power": "1.0", "train_snr_gamma": "1.0", "diff_alt_ratio": "1.0", "diff_revert_original_target": true, "diff_use_diff_mask": false, "model_v_pred": true, "mode": "ADDifT", }
Model Trimming: After various block weight xyz tests I used Supermerger to alter the model to :1.3:OUTALL 1.3 weight because I was using :2 in tests but the supermerger process would essentially triple the saved model weight if I kept it at 2.