Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation
# Shared defaults, not an additional training recipe.
model_name: Wan2.1-T2V-1.3B
load_diffusion_pretrained: true
generator_fsdp_sync_module_states: false
independent_first_frame: false
condition_only_first_frame: false
optimizer: adamw
lr_warmup_steps: 0
lr_drop_steps: []
lr_drop_gamma: 0.1
optimizer_config:
adamw:
lr: 1.0e-05
beta1: 0.0
beta2: 0.999
weight_decay: 0.01
i2v: false
gc_interval: 100
timestep_bucket_log_interval: 500
seed: 0
wds_sample_shuffle: 1000
wds_sample_shuffle_initial: 100
data_sources: null
inference_train_data_source_id: null
max_iters: -1
snapshot_iters: -1
checkpoint_path: null
resume_checkpoint_path: null
save_resume_checkpoint: false
batch_size: 1
val_batch_size: 1
test_sample_limit: -1
caption_dropout_p: 0.0
negative_prompt: Bright tones, overexposed, static, blurred details, subtitles, style,
works, paintings, images, static, overall gray, worst quality, low quality, JPEG
compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly
drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture,
messy background, three legs, many people in the background, walking backwards
max_grad_norm: 10.0
separate_predictive_grad_clip: false
max_predictive_grad_norm: null
max_checkpoints: 5
ema_checkpoint_mode: default
future_prediction:
enabled: false
init_mode: copy
source_context:
streams:
- clean
- noisy
source_layer_fusion:
mode: mlp
hidden_ratio: 2.0
activation: gelu
aux_future_timestep_range:
- 0.0
- 1000.0
model_kwargs:
timestep_shift: 5.0
local_attn_size: -1
sink_size: 0
timestep_shift: 5.0
sampling_steps: 50
sample_solver: unipc
val_num_workers: 4
no_save: false
no_visualize: false
disable_wandb: false
expdir: null
wandb_save_dir: null
wandb_host: https://api.wandb.ai
wandb_entity: null
wandb_project: Looking-Ahead
generator_ckpt: null
require_raw_generator_checkpoint: false
checkpoint_name: null
output_path: null
data_path: null
annotation_path: null
use_ema: false
use_train_data: false
duration: 5s
num_output_frames: -1
long_inference_threshold: 256
vae_decode_chunk: 32