Perflow-Shuai/Reproduce-Wan2.2-5B-CFG5-to-CFG1-50Step-LoRA-r64-iter500

Reproduce Wan2.2-TI2V-5B — CFG-only (native 50-step)

An independent GB300 training run: 500 iterations, 500 generator updates, LoRA rank/alpha 64/64, global batch 32, final topology 16 GPUs. This is a generator LoRA adapter and requires the base model.

Status: final checkpoint integrity/export checks and 16-case video generation are complete. Four-experiment video-quality acceptance against the published references remains pending. Publication does not claim equivalent quality or convergence; see the per-case observations and limitations in reference_comparison/review.json.

Files and use

• adapter_model.safetensors: final generator adapter; adapter_config.json specifies exact native Wan Linear targets.
• generator_lora.pt: equivalent native LongLive payload (generator_lora, step, adapter).
• training_config.yaml: training settings with portable model/data paths; inference_config.yaml: CFG=1, LoRA scale=1, UniPC shift5, 50 steps.
• frozen_source.tar: exact archived training source; the separately requested minimal code branch is still awaiting quality acceptance.
• publication_manifest.json and SHA256SUMS: export hashes. Full optimizer/critic/RNG continuation files remain on the training cluster.

Use the native Wan/LongLive wrappers from LongLive-LoRA, with the frozen source archive to recover the exact implementation. Download the base model into wan_models/Wan2.2-TI2V-5B (or update model_kwargs.model_dir), set checkpoints.lora_ckpt to the downloaded generator_lora.pt, and use the supplied inference settings. No negative/unconditional forward is needed at CFG=1. Generic Diffusers module names differ and may require key conversion; this is not a merged full model.

Data and comparison

Training text comes from the pinned requested prompt file. Exact revision, SHA-256 and split counts are in data_provenance.json. The unique texts were ordered by SHA-256 of reproduce-20260927: plus each prompt; the first16 were held out, the next64 selected for CFG teacher-trajectory caching. DMD uses the remaining complete training split; CFG-only uses the64-prompt cached-teacher recipe. The heldout texts are disjoint from this local training split; they are not proven unseen by the public reference.

The published reference and this adapter use matched prompts, seeds, FP32 initial noise, output dimensions and CFG=1. Comparison profile: 1280×704, 121 frames, 24fps. All16 public/local video pairs and contact sheets are under reference_comparison/. Training resolution can differ from comparison resolution; retain training_config.yaml as the training record. 14B DMD includes both published LightX2V Euler and common UniPC controls; all other comparisons use UniPC. The16 prompts form a diagnostic sample, not a broad equivalence benchmark.

Example local video: case00. Matching reference: case00.