task: t2v
work_dir: output/sana_video2_5b_720p
report_to: none
tracker_project_name: sana-video-2
data:
  data_dir:
    video: data/video_toy_data
  image_size: 720
  type: SanaZipDataset
  transform: default_train_video
  aspect_ratio_type: ASPECT_RATIO_VIDEO_720_MS_DIV32
  caption_selection_type: proportion
  caption_proportion:
    prompt: 1
  num_frames: 193
  target_fps: 24
  resample_fps: false
model:
  model: SanaVideo2_5B
  image_size: 720
  mixed_precision: bf16
  fp32_attention: true
  load_from: null
  multi_scale: false
  linear_head_dim: 128
  softmax_head_dim: 256
  softmax_ratio: 0.25
  ffn_type: SwiGLU
  mlp_ratio: 4.0
  use_pe: true
  pos_embed_type: wan_rope
  qk_norm: true
  cross_norm: true
  class_dropout_prob: 0.1
  attn_res_block_size: 8
vae:
  vae_type: LTX2VAE_diffusers
  vae_pretrained: output/pretrained_models/LTX-2.3-Diffusers
  weight_dtype: bfloat16
  vae_latent_dim: 128
  vae_downsample_rate: 32
  vae_stride: [8, 32, 32]
  sample_posterior: false
  use_causal_encode: true
text_encoder:
  text_encoder_name: gemma-2-2b-it
  caption_channels: 2304
  y_norm: true
  y_norm_scale_factor: 0.01
  model_max_length: 300
  chi_prompt:
    - 'Given a user prompt, generate an "Enhanced prompt" that provides detailed visual descriptions suitable for image generation. Evaluate the level of detail in the user prompt:'
    - '- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, and spatial relationships to create vivid and concrete scenes.'
    - '- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.'
    - 'Here are examples of how to transform or refine prompts:'
    - '- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.'
    - '- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus passing by towering glass skyscrapers.'
    - 'Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:'
    - 'User Prompt: '
scheduler:
  train_sampling_steps: 1000
  predict_flow_v: true
  noise_schedule: linear_flow
  pred_sigma: false
  learn_sigma: false
  flow_shift: 6.0
  inference_flow_shift: 12.0
  weighting_scheme: logit_normal
  logit_mean: 0.0
  logit_std: 0.95
  vis_sampler: flow_dpm-solver
train:
  use_fsdp: true
  num_workers: 8
  seed: 1
  train_batch_size: 1
  num_epochs: 10
  gradient_accumulation_steps: 1
  grad_checkpointing: true
  gradient_clip: 0.1
  joint_training_interval: 0
  optimizer:
    type: AdamW
    lr: 5.0e-5
    betas: [0.9, 0.999]
    eps: 1.0e-10
    weight_decay: 0.0
  auto_lr: null
  lr_schedule: constant
  lr_schedule_args:
    num_warmup_steps: 500
  log_interval: 10
  save_model_steps: 1000
  save_model_epochs: 1