ltxv_2b_lora.yaml

# LTXV LoRA Training Configuration

# Model configuration
model:
  model_source: "LTXV_2B_0.9.6_DEV" # Options: "LTXV_13B_097_DEV", "LTXV_2B_0.9.6_DEV", "LTXV_2B_0.9.5", "LTXV_2B_0.9.1", "LTXV_2B_0.9.0", or a HF repo/local path
  training_mode: "lora" # Options: "lora" or "full"
  load_checkpoint: null # Path to checkpoint file or directory to resume from. If directory, latest checkpoint will be used.

# LoRA configuration
lora:
  rank: 32
  alpha: 32
  dropout: 0.0
  target_modules:
    - "to_k"
    - "to_q"
    - "to_v"
    - "to_out.0"

# Conditioning configuration
conditioning:
  mode: "none" # Options: "none", "reference_video"
  first_frame_conditioning_p: 0.1

# Optimization configuration
optimization:
  learning_rate: 2e-4
  steps: 1500
  batch_size: 1
  gradient_accumulation_steps: 1
  max_grad_norm: 1.0
  optimizer_type: "adamw" # Options: "adamw" or "adamw8bit"
  scheduler_type: "linear" # Options: "constant", "linear", "cosine", "cosine_with_restarts", "polynomial"
  scheduler_params: {}
  enable_gradient_checkpointing: false

# Acceleration optimization
acceleration:
  mixed_precision_mode: "bf16" # Options: "no", "fp16", "bf16"
  quantization: null # Options: null, "int8-quanto", "int4-quanto", "int2-quanto", "fp8-quanto", "fp8uz-quanto"
  load_text_encoder_in_8bit: true # Load text encoder in 8-bit precision to save memory
  compile_with_inductor: false
  compilation_mode: "reduce-overhead" # Options: "default", "reduce-overhead", "max-autotune"

# Data configuration
data:
  preprocessed_data_root: "/path/to/preprocessed/data"
  num_dataloader_workers: 2

# Validation configuration
validation:
  prompts:
    - "a professional portrait video of a person with blurry bokeh background"
    - "a video of a person wearing a nice suit"
  negative_prompt: "worst quality, inconsistent motion, blurry, jittery, distorted"
  images: null # Set to a list of image paths to use first-frame conditioning, or null to disable
  video_dims: [768, 448, 89] # [width, height, frames]
  seed: 42
  inference_steps: 50
  interval: 100 # Set to null to disable validation
  videos_per_prompt: 1
  guidance_scale: 3.5

# Checkpoint configuration
checkpoints:
  interval: 250 # Save a checkpoint every N steps, set to null to disable
  keep_last_n: -1 # Keep only the N most recent checkpoints, set to -1 to keep all

# Flow matching configuration
flow_matching:
  timestep_sampling_mode: "shifted_logit_normal" # Options: "uniform", "shifted_logit_normal"
  timestep_sampling_params: {}

# HuggingFace Hub configuration
hub:
  push_to_hub: false # Whether to push the model weights to the Hugging Face Hub
  hub_model_id: null # Hugging Face Hub repository ID (e.g., 'username/repo-name'). Must be provided if `push_to_hub` is set to True

# W&B configuration
wandb:
  enabled: false  # Set to true to enable W&B logging
  project: "ltxv-trainer"
  entity: null  # Your W&B username or team
  tags: []
  log_validation_videos: true

# General configuration
seed: 42
output_dir: "outputs/ltxv_lora"