AnyFlow Continuation Quickstart¶

This guide is for continuing the AnyFlow training objective on a downstream Wan dataset. For the implementation overview, see AnyFlow.

The public NVIDIA AnyFlow checkpoints are full Diffusers pipelines with full transformer weights, not LoRA adapters. Do not point init_lora at those repositories. Use init_lora only when you have an actual SimpleTuner-compatible LoRA file or repository.

Which Checkpoint To Use¶

Use the bidirectional T2V AnyFlow checkpoints as the pretrained transformer:

• nvidia/AnyFlow-Wan2.1-T2V-1.3B-Diffusers
• nvidia/AnyFlow-Wan2.1-T2V-14B-Diffusers

Keep the original Wan checkpoint as the source for the text encoder, tokenizer, VAE, and scheduler:

• Wan-AI/Wan2.1-T2V-1.3B-Diffusers
• Wan-AI/Wan2.1-T2V-14B-Diffusers

The FAR checkpoints (nvidia/AnyFlow-FAR-*) use a causal AnyFlow transformer architecture and are not the target of this SimpleTuner quickstart.

Example Config¶

Start from the normal Wan quickstart config, then change the model and distillation fields:

{
  "model_family": "wan",
  "model_type": "lora",
  "pretrained_model_name_or_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
  "pretrained_t5_model_name_or_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
  "pretrained_vae_model_name_or_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
  "pretrained_transformer_model_name_or_path": "nvidia/AnyFlow-Wan2.1-T2V-1.3B-Diffusers",
  "pretrained_transformer_subfolder": "transformer",
  "data_backend_config": "config/wan/multidatabackend.json",
  "output_dir": "output/wan-anyflow-lora",
  "lora_rank": 32,
  "lora_alpha": 32,
  "train_batch_size": 1,
  "gradient_accumulation_steps": 1,
  "learning_rate": 0.0001,
  "max_train_steps": 1000,
  "distillation_method": "anyflow",
  "distillation_config": {
    "anyflow": {
      "target_mode": "online_teacher",
      "teacher_rollout_steps": 1,
      "r_timestep_sampler": "uniform",
      "min_interval_ratio": 0.02,
      "gate_value": 0.25,
      "deltatime_type": "r",
      "loss_weight": 1.0
    }
  }
}

Run training from the SimpleTuner directory:

simpletuner train

The resulting LoRA continues from the AnyFlow distilled transformer while keeping the AnyFlow objective active during downstream fine-tuning.

If You Have An AnyFlow LoRA¶

If an AnyFlow LoRA has been extracted and published separately, then use the original Wan base checkpoint and load the adapter with init_lora:

{
  "model_family": "wan",
  "model_type": "lora",
  "pretrained_model_name_or_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
  "pretrained_t5_model_name_or_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
  "pretrained_vae_model_name_or_path": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
  "init_lora": "your-org/anyflow-wan21-1.3b-lora",
  "lora_rank": 32,
  "lora_alpha": 32,
  "distillation_method": "anyflow"
}

The LoRA rank and target modules must match the published adapter. A full transformer checkpoint is not a valid init_lora value.

About Extracting A LoRA¶

Extracting a LoRA from a full AnyFlow transformer is possible in principle, but it is a conversion project rather than a training option. SimpleTuner includes experimental extraction scripts:

python scripts/extract_peft_lora.py \
  Wan-AI/Wan2.1-T2V-1.3B-Diffusers \
  nvidia/AnyFlow-Wan2.1-T2V-1.3B-Diffusers \
  output/anyflow-wan21-1.3b-r32.safetensors \
  --rank 32

For a LyCORIS/LoCon adapter, use scripts/extract_lycoris_adapter.py with the same arguments plus --algo locon.

The conversion does this:

1. Load the matching Wan base transformer and the AnyFlow transformer.
2. Diff matching linear-layer weights.
3. Factorize each delta into low-rank LoRA matrices.
4. Save a PEFT-compatible adapter with the exact target module list and rank.
5. Validate that generation and downstream continuation still behave like the full AnyFlow checkpoint.

This is approximate and rank-dependent. The default script target set matches SimpleTuner's Wan PEFT defaults (to_q,to_k,to_v,to_out.0). Use --target-modules all-linear only if the downstream config also targets the same modules.

Current Limits¶

• The public NVIDIA AnyFlow model license is noncommercial; check the upstream model card before publishing derived adapters.
• AnyFlow validation is wired through the distiller scheduler hook for registered FlowMap-capable pipelines. Custom or external validation paths still need to pass r_timestep or timestep_r into the model component.
• Full-rank online-teacher continuation still needs separate student and teacher wiring. LoRA continuation is the supported path for now.