Perflow-Shuai/Wan2.1-T2V-14B-CFG5-to-CFG1-50Step-LoRA-r128-iter500

Wan2.1-T2V-14B CFG-only distillation LoRA

This repository contains the iteration-500 generator LoRA from a CFG-only distillation run on Wan-AI/Wan2.1-T2V-14B. The training target is the base model's standard CFG=5 guided flow; the student is trained and evaluated at CFG=1. The repository name records the complete contract: CFG5-to-CFG1, native 50 steps, rank 128, and iteration 500.

This is not a few-step, DMD, SFP, or autoregressive checkpoint. It keeps the native 50-step schedule and removes the unconditional inference branch.

CFG scales and inference contract

Wan inference uses v_cfg = v_uncond + g * (v_cond - v_uncond). Therefore g=5 is standard CFG 5, while g=1 is exactly the conditional prediction and requires one transformer call per denoising step. CFG and LoRA weight scale are different controls.

Adapter contents

• —Generator-only LoRA; optimizer and training-state files are not published
• —800 finite FP32 tensors, 400 complete A/B pairs
• —613,416,960 LoRA parameters
• —400 Linear targets across all 40 transformer blocks
• —Per block: self-attention q/k/v/o, cross-attention q/k/v/o, and FFN 0/2
• —Native and Safetensors exports match the source checkpoint tensor-for-tensor

Files

• —adapter_model.safetensors: safe portable generator LoRA state dict
• —adapter_config.json: PEFT rank, alpha, and exact target-module names
• —generator_lora.pt: native LongLive payload with generator_lora and step=500
• —training_config.yaml: exact executed training configuration
• —inference_overrides.yaml: matched native-50-step CFG=1 settings
• —release_metadata.json: machine-readable training and evaluation contract
• —provenance.json: source hashes and export validation
• —SHA256SUMS: checksums for every published artifact

PyTorch pickle files should only be loaded from trusted sources. Prefer the Safetensors file when the target runtime supports it.

Download and LongLive usage

from huggingface_hub import hf_hub_download

lora_path = hf_hub_download(
    repo_id="Perflow-Shuai/Wan2.1-T2V-14B-CFG5-to-CFG1-50Step-LoRA-r128-iter500",
    filename="generator_lora.pt",
)
print(lora_path)

Configure the LongLive Wan2.1-T2V-14B 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: 128
  alpha: 128
  dropout: 0.0

Validation

The step-500 adapter was evaluated on eight held-out prompts using matched prompt, seed, initial noise, and scheduler state for three arms: Base CFG=5, Base CFG=1, and Distilled CFG=1 at LoRA scale 1. All 24 videos decoded with 81 unique frames at 832x480, 16 FPS. The auxiliary latent-distance gap ratio was below 1 for all eight cases (mean 0.7671, or 23.29% mean gap closure toward the CFG=5 reference). This is a small diagnostic suite, not a broad benchmark.

Training provenance and limitation

The run reached 500 cumulative optimizer updates at effective global batch 32. At step 92 it migrated from 16 data-parallel ranks with accumulation 2 to 8 ranks with accumulation 4. LoRA bytes, optimizer state, global sample order, and update boundaries were preserved, but FSDP reduction order and per-rank RNG streams changed; the continuation is sample-equivalent, not bitwise identical to an uninterrupted run.

The adapter is validated only for Wan2.1-T2V-14B with the matched 50-step, CFG=1, LoRA-scale-1 contract. Other schedulers, resolutions, lengths, CFG values, LoRA scales, and downstream fine-tunes are not established here.

Integrity

Original step-500 checkpoint SHA-256: af81887d941dcf6d0fb6c6925f7b38eecfbed5cfe444cc8e92988c10963d904f.

Training configuration SHA-256: f0ec25972040e12912f02704d69193085e47b7acc692173ad88339cb13ed4e2b.

Executed training code commit: 011a1d04a1f084fae1de1ee421da7f70efabf378. The public commit `b1d380ffaaad8706b6a9b4d4f3e80867db88a300` has the identical Git tree 7f576cd6ed371befa1a0afd408af100b55b63f99; the commit IDs differ only because the same change was replayed onto a different parent.

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