LTX-Video-Trainer/configs/ltxv_13b_ic_lora.yaml at main · Lightricks/LTX-Video-Trainer · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
# LTXV LoRA Training Configuration

# Model configuration
model:
  model_source: "LTXV_13B_097_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: 128
  alpha: 128
  dropout: 0.0
  target_modules:
    - "to_k"
    - "to_q"
    - "to_v"
    - "to_out.0"
    - "ff.net.0.proj"
    - "ff.net.2"

# Conditioning configuration
conditioning:
  mode: "reference_video" # Options: "none", "reference_video"
  first_frame_conditioning_p: 0.2
  reference_latents_dir: "reference_latents" # Directory for reference latents when using reference_video mode

# Optimization configuration
optimization:
  learning_rate: 2e-4
  steps: 2000
  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: true

# 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: false # 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:
    - "First validation prompt"
    - "Second validation prompt"
    - "Third validation prompt"
  reference_videos:
    - "/path/to/reference_a.mp4"
    - "/path/to/reference_b.mp4"
    - "/path/to/reference_c.mp4"
  negative_prompt: "worst quality, inconsistent motion, blurry, jittery, distorted"
  video_dims: [512, 512, 81] # [width, height, frames]
  seed: 42
  inference_steps: 50
  interval: 250 # 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: 3 # 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/ic_lora_13b"