Wan2.2-5B-NonAR-DMD-4Step-LoRA-r64-iter1600

Wan2.2-5B non-AR DMD LoRA — Rank 64 · Iter 1600 · 4-step inference

Checkpoint identity / 权重信息

Important: <strong>iter 1600</strong> is the training checkpoint iteration; <strong>4-step</strong> is the number of denoising steps used during inference.

This repository contains the inference-time generator LoRA shown in the <a href="https://competent-postposted-predictions-nikon.trycloudflare.com">pure Wan2.2-5B non-AR DMD rank-comparison page</a>. It is the rank-64 result at training iteration 1600, sampled in 4 denoising steps. The gallery compares rank 32/64/128/256 LoRAs against the official Wan2.2-TI2V-5B baseline; no ControlNet is used.

Checkpoint

• Base model: <a href="https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B">Wan-AI/Wan2.2-TI2V-5B</a>
• Objective: DMD with backward simulation
• Student / teacher / critic attention: full-sequence non-AR
• LoRA rank / alpha / scale: 64 / 64 / 1.0
• Training iteration: 1600 (<code>checkpoint_model_001600</code>)
• Training window: 32 latent frames
• Inference denoising steps: 4 (FlowUniPC, timestep shift 5)
• Targets: 300 Linear modules across 30 transformer blocks
• Per block: self-attention q/k/v/o, cross-attention q/k/v/o, FFN 0/2
• Published adapter: generator only; the training-only critic LoRA is intentionally omitted

The adapter contains 600 finite FP32 tensors (161,218,560 parameters), one LoRA A/B pair for each of the 300 listed target modules. Publication verifies that all tensors are finite and that the native and Safetensors files preserve the same generator keys and values. The pure-Wan gallery includes this rank-64/step-1600 checkpoint across eight validation prompts.

Files

• <code>adapter_model.safetensors</code>: generator-only PEFT adapter, safe portable format
• <code>generator_lora.pt</code>: generator-only native LongLive checkpoint with keys <code>generator_lora</code> and <code>step</code>
• <code>adapter_config.json</code>: rank, alpha, and the exact 300 target-module names
• <code>training_config.yaml</code>: source non-AR DMD training configuration
• <code>inference_overrides.yaml</code>: essential settings for 4-step LongLive inference
• <code>provenance.json</code>: source and published-file checksums

LongLive usage

Download the native checkpoint:

from huggingface_hub import hf_hub_download

lora_path = hf_hub_download(
    repo_id="Perflow-Shuai/Wan2.2-5B-NonAR-DMD-4Step-LoRA-r64-iter1600",
    filename="generator_lora.pt",
)
print(lora_path)

In the LongLive inference config, keep the normal Wan2.2-TI2V-5B base model and apply these settings:

algorithm:
  generator_is_causal: false

model_kwargs:
  model_name: Wan2.2-TI2V-5B
  timestep_shift: 5.0
  num_frame_per_block: 8
  local_attn_size: -1

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

inference:
  sampling_steps: 4
  sink_size: 0
  multi_shot_rope_offset: 8

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

The current LongLive loader recognizes <code>generator_lora.pt</code> directly. For other Wan runtimes, use <code>adapter_model.safetensors</code> and map the exact target names from <code>adapter_config.json</code>. The comparison page tested repository-specific CFG values for controlled comparisons; those values should not be treated as a universal recommended CFG recipe.

Provenance and integrity

The original training checkpoint contained equal-size generator and critic LoRAs plus <code>step=1600</code>. Publication extracts the generator tensors without changing their names, shapes, dtypes, or values.

Original combined training-checkpoint SHA-256: <code>f4dfbc89ea70351213ed3b153f731025f572705eea36830e6a71f6cd2c2c2115</code>.