config/examples/train_lora_wan22_14b_24gb.yaml

# this example focuses mainly for training Wan2.2 14b on images. It will work for video as well by increasing
# the number of frames in the dataset and samples. Training on and generating video is very VRAM intensive.
---
job: extension
config:
  # this name will be the folder and filename name
  name: "my_first_wan22_14b_lora_v1"
  process:
    - type: 'sd_trainer'
      # root folder to save training sessions/samples/weights
      training_folder: "output"
      # uncomment to see performance stats in the terminal every N steps
#      performance_log_every: 1000
      device: cuda:0
      # Use a trigger word if train.unload_text_encoder is true, however, if caching text embeddings, do not use a trigger word
      # trigger_word: "p3r5on"
      network:
        type: "lora"
        linear: 32
        linear_alpha: 32
      save:
        dtype: float16 # precision to save
        save_every: 250 # save every this many steps
        max_step_saves_to_keep: 4 # how many intermittent saves to keep
      datasets:
        # datasets are a folder of images. captions need to be txt files with the same name as the image
        # for instance image2.jpg and image2.txt.
        # "C:\\path\\to\\images\\folder"
        - folder_path: "/path/to/images/or/video/folder"
          caption_ext: "txt"
          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
          # number of frames to extract from your video. It will automatically extract them evenly spaced
          # set to 1 frame for images
          num_frames: 1
          resolution: [ 512, 768, 1024]
      train:
        batch_size: 1
        steps: 2000  # total number of steps to train 500 - 4000 is a good range
        gradient_accumulation: 1
        train_unet: true
        train_text_encoder: false  # probably won't work with wan
        gradient_checkpointing: true  # need the on unless you have a ton of vram
        noise_scheduler: "flowmatch" # for training only
        timestep_type: 'linear'
        optimizer: "adamw8bit"
        lr: 1e-4
        optimizer_params:
          weight_decay: 1e-4
        # uncomment this to skip the pre training sample
#        skip_first_sample: true
        # uncomment to completely disable sampling
#        disable_sampling: true
        dtype: bf16

        # IMPORTANT: this is for Wan 2.2 MOE. It will switch training one stage or the other every this many steps
        switch_boundary_every: 10

        # required for 24GB cards. You must do either unload_text_encoder or cache_text_embeddings but not both

        # this will encode your trigger word and use those embeddings for every image in the dataset, captions will be ignored
        # unload_text_encoder: true

        # this will cache all captions in your dataset.
        cache_text_embeddings: true

      model:
        # huggingface model name or path, this one if bf16, vs the float32 of the official repo
        name_or_path: "ai-toolkit/Wan2.2-T2V-A14B-Diffusers-bf16"
        arch: 'wan22_14b'
        quantize: true
        # This will pull and use a custom Accuracy Recovery Adapter to train at 4bit
        qtype: "uint4|ostris/accuracy_recovery_adapters/wan22_14b_t2i_torchao_uint4.safetensors"
        quantize_te: true
        qtype_te: "qfloat8"
        low_vram: true
        model_kwargs:
          # you can train high noise, low noise, or both. With low vram it will automatically unload the one not being trained.
          train_high_noise: true
          train_low_noise: true
      sample:
        sampler: "flowmatch"
        sample_every: 250 # sample every this many steps
        width: 1024
        height: 1024
        # set to 1 for images
        num_frames: 1
        fps: 16
        # samples take a long time. so use them sparingly
        # samples will be animated webp files, if you don't see them animated, open in a browser.
        prompts:
          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
#          - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
          - "woman with red hair, playing chess at the park, bomb going off in the background"
          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
          - "a bear building a log cabin in the snow covered mountains"
          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
          - "hipster man with a beard, building a chair, in a wood shop"
          - "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"
          - "a man holding a sign that says, 'this is a sign'"
          - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
        neg: ""  
        seed: 42
        walk_seed: true
        guidance_scale: 3.5
        sample_steps: 25
# you can add any additional meta info here. [name] is replaced with config name at top
meta:
  name: "[name]"
  version: '1.0'