This repository contains a BF16 LoRA adapter for causal chunk-autoregressive inference with Wan2.1-T2V-14B. It is an intermediate Stage-1 checkpoint from a four-stage Resampling Forcing training campaign.

Training status: step 4,585 of 31,500. This checkpoint has only received the causal teacher-forcing warm-up objective. Resampling Forcing starts at step 10,000, so this release is not a completed or RF-trained reproduction.

Checkpoint identity

Item	Value
Base model	`Wan-AI/Wan2.1-T2V-14B`
Base revision	`a064a6c71f5be440641209c07bf2a5ce7a2ff5e4`
Training step	4,585 of 31,500
Training samples	293,440 at global batch 64
Current objective	Stage 1 causal teacher forcing
LoRA rank / alpha / dropout	32 / 32 / 0.0
Adapter dtype	BF16
Adapter parameters	153,354,240 (1.0619% of wrapped model)
Adapter tensors	800 tensors, 400 A/B pairs

The adapter targets every Linear layer inside each of the 40 causal Wan transformer blocks: self-attention q/k/v/o, cross-attention q/k/v/o, and FFN projections 0/2. The base model weights are not included.

AR inference profile

Causal chunk size: 3 latent frames
5-second profile: 21 latent frames in 7 chunks
Output: 81 RGB frames, 832x480, 16 fps
Dense causal history with clean K/V recache per chunk
Euler, 32 denoising steps per chunk, CFG 5.0, timestep shift 5.0
LoRA scale: 1.0
Conditional and unconditional CFG branches use separate caches

This adapter requires the custom causal Wan implementation in the shyang/resampling_forcing_14b branch. It is not a standalone model and is not guaranteed to load in stock Diffusers.

bash

python scripts/infer_resampling_forcing.py \
  --config-path configs/resampling_forcing/wan21_14b_long/stage1_tf_5s.yaml \
  --checkpoint /path/to/adapter_model.safetensors \
  --prompt "A cinematic continuous shot of ..." \
  --duration-seconds 5 \
  --output-dir outputs/example \
  --seed 0 \
  --dtype bf16 \
  --cfg-devices cuda:0 cuda:1

Planned training schedule

Stage	Objective	Profile	Steps
1	Causal teacher forcing	5 s, dense history	0-10,000
2	Resampling Forcing	5 s, dense history	10,000-25,000
3	Resampling Forcing	15 s, dense history	25,000-30,000
4	Resampling Forcing	15 s, top-5 history routing	30,000-31,500

Validation status

The SafeTensors export is checked tensor-by-tensor against the step-4,585 training checkpoint. All 800 tensors and 153,354,240 parameters matched, and the exported file passed the project inference loader at commit 5c926d1. The existing same-prompt visual AR comparison was generated from the earlier step-4,503 checkpoint, not this exact step-4,585 artifact. No quantitative generation benchmark or safety evaluation has been completed for this intermediate release.

Files

adapter_model.safetensors: generator-only BF16 LoRA for inference.
adapter_config.json: adapter identity and PEFT metadata.
inference_config.json: validated 5-second causal AR inference profile.
training_config.yaml: source Stage-1 training configuration.
training_state/step_004585/model.pt: exact trainer state with BF16 LoRA, AdamW state, and RNG states.
export_adapter.py: exporter used to create the SafeTensors artifact.
provenance.json and SHA256SUMS: revisions, sizes, and integrity hashes.

The file under training_state/ is a PyTorch pickle checkpoint. Load it only in a trusted environment. Use adapter_model.safetensors for inference.

File	Bytes	SHA-256
`adapter_model.safetensors`	306,809,056	`ababb2ec20470b0f04d62f47e503b2776928b587f752028fb822424edba8b74a`
`training_state/step_004585/model.pt`	921,650,933	`41c045fb8778f628cda613a3bfa522df75e2613b5655954467f967adb733b1c7`

Limitations and responsible use

This intermediate research checkpoint may produce temporal drift, visual artifacts, prompt failures, or unsafe content and inherits the limitations of Wan2.1. Do not use it for high-stakes or deceptive applications. Training-data rights are not conveyed by this repository.

License and attribution

The adapter and project code are distributed under Apache-2.0. The Wan2.1-T2V-14B base model is also published under Apache-2.0 and must be obtained separately. This work implements the schedule described in the Resampling Forcing paper.

Perflow-Shuai/Wan2.1-T2V-14B-Resampling-Forcing-AR-LoRA

Wan2.1-T2V-14B Resampling-Forcing AR LoRA - Stage-1 Step 4585