Perflow-Shuai/SCOPE-Wan2.2-5B-NonAR-DMD-4Step-LoRA-r32-iter2000

SCOPE Wan2.2-5B non-AR DMD LoRA — Rank 32 · Iter 2000 · 4-step inference

Checkpoint identity / 权重信息

Item	Value
LoRA rank / alpha	32 / 32
Training iteration / 训练迭代	2000 (checkpoint_model_002000)
Inference denoising steps / 推理步数	4
Required base model	zizhaotong/SCOPE
Foundation model	Wan-AI/Wan2.2-TI2V-5B
Architecture	full-sequence bidirectional non-AR
Conditioning	first image + per-frame 10-DoF action + text context

Important: iter 2000 is the training checkpoint iteration; 4-step is the number of denoising steps used during inference.

This repository contains the inference-time generator LoRA distilled on top of the released SCOPE action-conditioned world model. Apply it to SCOPE, not to vanilla Wan2.2-TI2V-5B. Although the 300 adapted base-DiT projections are structurally shared with vanilla Wan2.2, such cross-base use is out of the training contract.

Training and adapter details

• Objective: DMD with backward simulation
• Student / teacher / critic attention: full-sequence, bidirectional non-AR
• SCOPE conditioning: first frame fixed; 10-DoF action conditioning enabled
• LoRA rank / alpha / scale: 32 / 32 / 1.0
• Training iteration: 2000 (checkpoint_model_002000)
• Training pool: 66,521 prepared CrossFPS records, one manifest pass
• Training window: 21 latent frames / 81 raw action frames at 20 FPS
• Inference solver: 4-step 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
• ActionModule LoRA targets: none; the released SCOPE ActionModules remain active
• Published adapter: generator only; critic LoRA and both Adam states are omitted

The prepared source records are normalized 5-second clips with 100 action rows. This training run intentionally retains the requested 21-latent setup, so the loader uses action rows 0–80. The generated video contains 81 frames (4.05 seconds as a file at 20 FPS; 4.0 seconds between first and last frame), not the complete 5-second source horizon.

The adapter contains 600 finite BF16 tensors and 80,609,280 parameters: one LoRA A/B pair for each of the 300 target modules. Publication checks that the native and Safetensors exports have identical keys, shapes, dtypes, and tensor values.

Files

• adapter_model.safetensors: generator-only PEFT adapter in safe portable format
• generator_lora.pt: native LongLive checkpoint with generator_lora and step
• adapter_config.json: rank, alpha, and the exact 300 target-module names
• training_config.yaml: the actual full-data training configuration
• inference_overrides.yaml: essential SCOPE-aware 4-step inference settings
• provenance.json: source identity, dataset/config identity, and file checksums

LongLive SCOPE usage

Download the native checkpoint:

from huggingface_hub import hf_hub_download

lora_path = hf_hub_download(
    repo_id="Perflow-Shuai/SCOPE-Wan2.2-5B-NonAR-DMD-4Step-LoRA-r32-iter2000",
    filename="generator_lora.pt",
)
print(lora_path)

With the SCOPE-aware LongLive inference entry point, run the official example_0 input as follows:

python inference_scope.py \
  --model-dir /path/to/SCOPE \
  --image /path/to/SCOPE-code/examples/example_0/image.png \
  --action /path/to/SCOPE-code/examples/example_0/action.parquet \
  --prompt-file /path/to/SCOPE-code/examples/example_0/prompt.txt \
  --lora-checkpoint /path/to/generator_lora.pt \
  --lora-rank 32 \
  --lora-alpha 32 \
  --steps 4 \
  --solver unipc \
  --num-frames 81 \
  --fps 20 \
  --output scope_dmd_4step.mp4

The upstream SCOPE CLI does not currently expose a LoRA checkpoint argument; use a SCOPE-aware PEFT integration. For other runtimes, load adapter_model.safetensors with the exact module names in adapter_config.json. Keep the SCOPE ActionModules and image/action inputs enabled.

Provenance and integrity

The original training checkpoint also contains a same-size critic LoRA and two optimizer states. Publication extracts only generator_lora without changing tensor names, shapes, dtypes, or values.

File	Bytes	SHA-256
adapter_model.safetensors	161,293,448	05ef8172b178df61f269a3c5c2c352b5f8e39651b8ac6a1ae1f6ffdd0d0537b3
generator_lora.pt	161,434,310	470912fffa8fdc0155fde04dd93ca42cfc02063af72e8f247f485bf35bb1a8e2

Original combined training checkpoint:

• Size: 968,842,882 bytes
• SHA-256: 2abbdc102bd5c30ee6b7e20374623396a9a64334213306afb5fdf3e57b5874e3

The uploaded training_config.yaml has SHA-256 645ec9cdb8cf8806004d26658574e196c2a532e20563a6fbd9aa09c68d43161b. Training used repository HEAD 655718bd6b27950b20c269b2c3275d786cc75348 plus local SCOPE integration changes, so that commit alone is not claimed as a complete reproduction of the run. See provenance.json for the full audit.