Shoulder deep fisting [Anima]

v1.0

Twin sister of my already existing LoRA, trained on same dataset as latest version. Training params is as follows:

  "network_dim": 20, 
  "network_alpha": 20, #treat dim and alpha as one param. Their individual sizes influence amount of details able to fit inside of lora and their ratio alpha/dim is specifing where your lora will be most active during generation steps (alpha/dim<1 more details, alpha/dim>1 more color blobs on earlier stages)
  "learning_rate": 0.0001,
  "max_train_epochs": 10,
  "resolution": 768,
  "repeats": 10,
  "caption_dropout": 0.1,
  "gpu_index": "0", #you won't be able to use gpu without correct Torch+CUDA pair
  "optimizer_type": "AdamW8bit",
  "lr_scheduler": "cosine_with_restarts",
  "lr_scheduler_num_cycles": 1,
  "lr_warmup_steps": 100,
  "train_batch_size": 1,
  "gradient_accumulation_steps": 1,
  "max_grad_norm": 1.0,
  "save_every_n_epochs": 1,
  "save_last_n_epochs": 4,
  "mixed_precision": "bf16",
  "gradient_checkpointing": true,
  "seed": 42,
  "noise_offset": 0.03,
  "multires_noise_discount": 0.3,
  "timestep_sampling": "sigmoid",
  "discrete_flow_shift": 1.0,

Base
Apparently, newly emmerged Anima model is a completly different architecture from sd or sdxl, making it mostly incompatible with many good loras. However, training loras for it is not as hard, as one may think. Standalone trainer app from citron is able to make aroud 5000 steps in 3 hours and is using around 5gb of VRAM. It is not as smooth in installation process as one may wish for, but still managable. Especially if you know/can bother to learn how to correctly install PyTorch for your specific version of CUDA driver of GPU. I myself figured it out just from couple of searches in google, so nothing it is not a rocket science.