Wan2.2-T2V-A14B-Diffusers

Wan 2.2 T2V-A14B is the successor to Wan 2.1, also a text-to-video flow-matching DiT. Its defining feature is a two-stage denoising pipeline: a high-noise transformer handles the early/noisy timesteps and a low-noise transformer_2 handles the later/cleaner timesteps, switching at boundary_ratio * num_train_timesteps (default 0.875). Each transformer is ~14B parameters, for ~28B total.

Set up NeMo AutoModel with the latest container or follow the installation instructions.

Fine-Tune Wan2.2-T2V-A14B-Diffusers

From the repository root, run:

uv run torchrun --nproc-per-node=8 \
  examples/diffusion/finetune/finetune.py \
  -c examples/diffusion/finetune/wan2_2_t2v_flow.yaml \
  --model.stage=high_noise \
  --fsdp.cpu_offload=true

Choose a Workflow

Goal	Start Here
Fine-tune - two-stage with `model.stage` knob	Use wan2_2_t2v_flow.yaml.
Inference - loads both stage checkpoints	Use generate_wan22.yaml.

Two-stage finetuning workflow

Because each transformer is ~14B parameters, NeMo AutoModel finetunes them one at a time:

Preprocess once - produces a single cached .meta set reusable across both stages:

python -m tools.diffusion.preprocessing_multiprocess video \
    --video_dir /path/to/videos --output_dir /path/to/wan22_cache \
    --processor wan2.2 --caption_format meta_json --caption_field caption \
    --resolution_preset 512p --target_frames 81

Finetune the high-noise stage (pipe.transformer, sigma in [boundary_ratio, 1.0]):

torchrun --nproc-per-node=8 \
    examples/diffusion/finetune/finetune.py \
    -c examples/diffusion/finetune/wan2_2_t2v_flow.yaml \
    --model.stage=high_noise \
    --data.dataloader.cache_dir=/path/to/wan22_cache \
    --checkpoint.checkpoint_dir=./WAN22_CKPT/wan22_high \
    --fsdp.cpu_offload=true

Finetune the low-noise stage (pipe.transformer_2, sigma in [0.0, boundary_ratio]):

torchrun --nproc-per-node=8 \
    examples/diffusion/finetune/finetune.py \
    -c examples/diffusion/finetune/wan2_2_t2v_flow.yaml \
    --model.stage=low_noise \
    --data.dataloader.cache_dir=/path/to/wan22_cache \
    --checkpoint.checkpoint_dir=./WAN22_CKPT/wan22_low \
    --fsdp.cpu_offload=true

Run inference loading both stage checkpoints:

python examples/diffusion/generate/generate.py \
    -c examples/diffusion/generate/configs/generate_wan22.yaml

Each finetuning run only holds one of the two transformers on GPU - the recipe drops the unused one before sharding so an FSDP2 dp=8 setup on 8x80GB H100 fits a single 14B model plus its AdamW state. --fsdp.cpu_offload=true is recommended; it moves the sharded params and optimizer state to host RAM during the step boundary.

Model Reference

Model Architecture

Property	Value
Task	Text-to-Video
Architecture	DiT (Flow Matching), two-stage
Parameters	14B + 14B (high-noise + low-noise)
Hugging Face Organization	Wan-AI

Task

Text-to-Video (T2V)

Available Models

Model	HF ID
Wan 2.2 T2V-A14B	`Wan-AI/Wan2.2-T2V-A14B-Diffusers`