---
license: apache-2.0
base_model: Wan-AI/Wan2.1-T2V-14B
library_name: peft
pipeline_tag: text-to-video
model_name: Wan2.1 T2V 14B CFG5-to-CFG1 native-50-step LoRA r128 iter500
tags:
  - wan2.1
  - wan2.1-t2v-14b
  - text-to-video
  - lora
  - cfg-distillation
  - native-step
  - rank128
  - 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](https://huggingface.co/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

| Setting | Value |
| --- | ---: |
| Teacher CFG used for distillation | **5.0** |
| Student CFG during training | **1.0** |
| Recommended inference CFG | **1.0** |
| Default LoRA weight scale | **1.0** |
| Teacher to student denoising steps | **50 to 50** |
| Scheduler / timestep shift | FlowUniPC / **5.0** |
| LoRA rank / alpha / dropout | **128 / 128 / 0** |
| Training iteration | **500** |

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

```python
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:

```yaml
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

| Weight file | Bytes | SHA-256 |
| --- | ---: | --- |
| `generator_lora.pt` | 2453955566 | `0ceb9d5f8a16e413900e59c8643465da896cafcb4eb2d42481423cd666115344` |
| `adapter_model.safetensors` | 2453769728 | `1a311f6030a74e739d9705347079a131bd5a5579bd693c993026d63e36d4bea0` |

Original step-500 checkpoint SHA-256:
`af81887d941dcf6d0fb6c6925f7b38eecfbed5cfe444cc8e92988c10963d904f`.

Training configuration SHA-256: `f0ec25972040e12912f02704d69193085e47b7acc692173ad88339cb13ed4e2b`.

Executed training code commit: `011a1d04a1f084fae1de1ee421da7f70efabf378`. The public commit
[`b1d380ffaaad8706b6a9b4d4f3e80867db88a300`](https://github.com/AndysonYs/LongLive-LoRA/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.