Wan2.2-5B-CFG5-to-CFG1-50Step-LoRA-r64-iter250

Wan2.2-5B CFG-only distillation LoRA

This is the generator LoRA from a CFG-only distillation run on Wan-AI/Wan2.2-TI2V-5B. It distills the base model's standard CFG=5 guided flow into one conditional student forward at CFG=1. It does not distill the denoising schedule and is not a few-step/DMD checkpoint.

CFG scales and inference contract

At inference, guidance_scale: 1.0 means conditional-only denoising: one DiT forward per denoising step and no negative/unconditional branch. CFG scale and LoRA weight scale are different controls; the training-aligned LoRA weight scale is 1.0.

What was distilled

The frozen teacher target at each cached native-trajectory state is

v_cfg5 = v_uncond + 5 * (v_cond - v_uncond)

The trainable base+LoRA student makes one conditional forward and minimizes relative-guidance-weighted MSE to v_cfg5. The run uses frozen teacher trajectory states; the student does not generate its own rollout during training (no self-forcing/self-generation loop). There is no DMD critic.

The run keeps the native 50-step schedule. Its training configuration is included verbatim as training_config.yaml.

Adapter contents

• 600 FP32 tensors, 161,218,560 LoRA parameters
• 300 Linear targets across all 30 transformer blocks
• Per block: self-attention q/k/v/o, cross-attention q/k/v/o, FFN 0/2
• Generator adapter only; no optimizer or critic state is published
• Native and Safetensors exports were checked tensor-by-tensor for exact equality

Files

• adapter_model.safetensors: generator LoRA in safe portable format
• adapter_config.json: PEFT configuration and exact 300 target modules
• generator_lora.pt: native LongLive payload with generator_lora and step
• training_config.yaml: exact training configuration
• inference_overrides.yaml: native-50-step, CFG=1 inference settings
• release_metadata.json: machine-readable CFG and evaluation contract
• provenance.json: source identities and local file checksums
• SHA256SUMS: checksums for the published artifacts

Download and LongLive usage

from huggingface_hub import hf_hub_download

lora_path = hf_hub_download(
    repo_id="Perflow-Shuai/Wan2.2-5B-CFG5-to-CFG1-50Step-LoRA-r64-iter250",
    filename="generator_lora.pt",
)
print(lora_path)

Configure the LongLive Wan2.2-5B inference path with:

checkpoints:
  lora_ckpt: /path/to/generator_lora.pt

inference:
  sampling_steps: 50
  guidance_scale: 1.0
  lora_weight_scale: 1.0

adapter:
  type: lora
  rank: 64
  alpha: 64
  dropout: 0.0

The matched evaluation profile used 1280×704 output, 121 frames at 24 FPS, 50 FlowUniPC steps, base default CFG=5, and distilled inference CFG=1.

Integrity and provenance

Original combined step-250 checkpoint SHA-256: ca079c126dda1744436174001cf419ce4d0ffca23b419c01f2d63a5155c1598b.

Training configuration SHA-256: 72060c553a8137df63c1dee4e3d657abc7d192c9351dfe405ef36995be23290c.

Training code commit: 4b33b3aae331c7811ed51374532a734df919ffcb.

This is a research checkpoint. The release records artifact integrity and the executed training contract; it does not claim universal prompt or downstream model quality.