questro_ltx_full / training_config.yaml
TheMindExpansionNetwork's picture
Upload folder using huggingface_hub
1e75ede verified
acceleration:
load_text_encoder_in_8bit: false
mixed_precision_mode: bf16
quantization: null
checkpoints:
interval: 300
keep_last_n: 10
data:
num_dataloader_workers: 2
preprocessed_data_root: /workspace/data/raw/.precomputed
flow_matching:
timestep_sampling_mode: shifted_logit_normal
hub:
hub_model_id: TheMindExpansionNetwork/questro_LTX
push_to_hub: true
lora:
alpha: 32
dropout: 0
rank: 32
target_modules:
- attn1.to_k
- attn1.to_q
- attn1.to_v
- attn1.to_out.0
- attn2.to_k
- attn2.to_q
- attn2.to_v
- attn2.to_out.0
model:
model_path: /workspace/models/ltx2/ltx-2-19b-dev.safetensors
text_encoder_path: /workspace/models/gemma
training_mode: lora
optimization:
batch_size: 1
enable_gradient_checkpointing: true
gradient_accumulation_steps: 1
learning_rate: 0.0001
max_grad_norm: 1.0
optimizer_type: adamw
scheduler_params:
end_factor: 0.1
start_factor: 1.0
scheduler_type: linear
steps: 2500
output_dir: /workspace/outputs/training_run
seed: 42
training_strategy:
first_frame_conditioning_p: 0.1
name: text_to_video
with_audio: false
validation:
frame_rate: 1
generate_audio: false
guidance_scale: 4
images: null
include_reference_in_output: false
inference_steps: 20
interval: 100
prompts:
- Questro is walking in a dystopian universe
reference_videos: null
seed: 42
skip_initial_validation: false
stg_blocks:
- 29
stg_mode: stg_v
stg_scale: 1
video_dims:
- 768
- 768
- 1
videos_per_prompt: 1
wandb:
enabled: false
entity: null
log_validation_videos: false
project: ltx2-training
tags: []