ltx2_tsantsalize_h100.yaml

# tsantsalize IC-LoRA (Phase 3). Same recipe as the other two. 98 pairs (49 img + 49 vid@97f).
# Buckets reflect this dataset: images 768x1024/1024x768/1440x1920, videos 768x1024x97/1024x768x97.
model:
  model_path: /workspace/models/ltx-2.3-22b-dev.safetensors
  text_encoder_path: /workspace/models/gemma-3-12b-it/
  training_mode: lora
  load_checkpoint: /workspace/outputs/ltx2_tsantsalize/checkpoints  # fresh -> step 0

lora:
  rank: 64
  alpha: 64
  dropout: 0.0
  target_modules:
    - attn1.to_q
    - attn1.to_k
    - attn1.to_v
    - attn1.to_out.0
    - attn2.to_q
    - attn2.to_k
    - attn2.to_v
    - attn2.to_out.0
    - ff.net.0.proj
    - ff.net.2

training_strategy:
  name: video_to_video
  reference_latents_dir: reference_latents
  first_frame_conditioning_p: 0.0

data:
  preprocessed_data_root: /workspace/dataset_tsantsalize/.precomputed
  num_dataloader_workers: 4

optimization:
  batch_size: 1
  gradient_accumulation_steps: 4
  enable_gradient_checkpointing: true    # REQUIRED for 22B even on 94GB
  max_grad_norm: 1.0
  optimizer_type: prodigy
  optimizer_params:
    decouple: true
    d_coef: 1.0
    safeguard_warmup: true
    use_bias_correction: true
    weight_decay: 0.01
  learning_rate: 1.0
  scheduler_type: constant
  scheduler_params: {}
  steps: 3000

acceleration:
  mixed_precision_mode: bf16
  quantization: null
  load_text_encoder_in_8bit: false

flow_matching:
  timestep_sampling_mode: shifted_logit_normal
  timestep_sampling_params: {}

validation:
  interval: 1000
  skip_initial_validation: true
  inference_steps: 30
  guidance_scale: 4.0
  negative_prompt: "worst quality, blurry, distorted, low quality, artifacts, deformed face, asymmetric features, generic face"
  prompts:
    - tsantsalize
    - tsantsalize
    - tsantsalize
  reference_videos:
    - /workspace/dataset_tsantsalize/reference/pair_00005.png   # natural input (768x1024)
    - /workspace/dataset_tsantsalize/reference/pair_00007.png   # natural input (768x1024)
    - /workspace/dataset_tsantsalize/reference/pair_00008.png   # natural input (768x1024)
  reference_downscale_factor: 1
  video_dims:                  # 81-frame video val (was F=1 image); 576x768 keeps 0.75 AR + safe VRAM
    - 576
    - 768
    - 81
  videos_per_prompt: 1
  frame_rate: 24.0
  include_reference_in_output: true
  generate_audio: false
  seed: 42
  stg_blocks:
    - 29
  stg_mode: stg_v
  stg_scale: 1.0

checkpoints:
  interval: 500
  keep_last_n: -1
  keep_training_states_last_n: 1
  save_training_state: full
  no_resume: false
  precision: bfloat16
  filename_prefix: tsantsalize

wandb:
  enabled: true
  entity: null
  project: ltx2-tsantsalize
  tags:
    - ltx2
    - ic-lora
    - tsantsalize
    - image+video
    - h100-fullres
  log_validation_videos: true

output_dir: /workspace/outputs/ltx2_tsantsalize
seed: 42