Wan2.1 T2V 1.3B Non-AR DMD 4-Step LoRA
This repository contains the iteration-1,600 student/generator LoRA from
an experimental Wan2.1 text-to-video DMD run.
This is a non-autoregressive (non-AR) DMD checkpoint. The student
generator is configured as bidirectional, not causal or blockwise
autoregressive. The training implementation uses the Self-Forcing-Plus DMD
loss, sampler, and update order, but the student rollout itself is non-AR.
Model roles
| Role | Model | Training state |
|---|
| Student / generator | Wan-AI/Wan2.1-T2V-1.3B | Rank-64 LoRA |
| Fake-score critic | Wan-AI/Wan2.1-T2V-1.3B | Separate rank-64 LoRA |
| Real-score teacher | Wan-AI/Wan2.1-T2V-14B | Frozen, no LoRA |
The 14B teacher is used only during DMD training. Inference requires the
Wan2.1-T2V-1.3B base model and the published generator LoRA; it does not
require the 14B teacher or the critic.
Checkpoint identity
| Item | Value |
|---|
| Training iteration | 1,600 of 3,000 |
| Generator type | Bidirectional / non-AR |
| Inference denoising steps | 4 |
| Denoising timesteps | 1000, 750, 500, 250 |
| LoRA rank / alpha / dropout | 64 / 64 / 0.0 |
| LoRA targets | 300 linear modules |
| Published LoRA | 600 BF16 tensors, 300 A/B pairs |
| Trainable parameters per adapter | 87,490,560 |
| Training latent frames | 21 |
| Decoded inference frames | 81 at 16 fps |
| External inference CFG | Disabled; guidance is distilled |
This is an intermediate checkpoint from a run that completed 3,000
iterations. The published Safetensors file contains only the student
generator adapter. The separate critic adapter and optimizer state are not
included.
Training configuration
| Parameter | Value |
|---|
| Hardware | 64 H100 GPUs, 8 nodes |
| Per-rank / global batch size | 1 / 64 |
| Generator learning rate | 1e-5 |
| Critic learning rate | 2e-6 |
| Generator / critic Adam beta1, beta2 | 0.0, 0.999 |
| Critic : generator update ratio | 5 : 1 |
| Distribution loss | DMD |
| Timestep shift | 5.0 |
| Training guidance scale | 4.0 |
| EMA | Disabled for LoRA training |
At step 1,600, the validated Adam counters are 320 generator updates and
1,600 critic updates.
Files
generator_lora_lightx2v.safetensors: generator-only BF16 rank-64 LoRA.inference_config.json: tested LightX2V four-step inference settings.training_config.yaml: source DMD training configuration.export_generator_lora.py: exporter used to create the inference adapter.provenance.json: revisions, role definitions, validation facts, and hash.SHA256SUMS: checksum for the published adapter.
LightX2V inference
Set lora_configs[0].path in inference_config.json to the absolute path of
generator_lora_lightx2v.safetensors, then run from a LightX2V checkout:
1python -m torch.distributed.run --nproc_per_node=8 -m lightx2v.infer \
2 --model_cls wan2.1_distill \
3 --task t2v \
4 --model_path /path/to/Wan2.1-T2V-1.3B \
5 --config_json /path/to/inference_config.json \
6 --prompt "your prompt" \
7 --negative_prompt "your negative prompt" \
8 --save_result_path output.mp4 \
9 --seed 42
The supplied configuration uses LoRA strength 1.0, four denoising steps,
81 output frames, 832x480 resolution, and no external CFG pass.
Integrity
| File | Bytes | SHA-256 |
|---|
generator_lora_lightx2v.safetensors | 175,045,992 | 73f0750f12f5f91a913960c77cc002f0256dc88133d1638dd0a3d491277d7866 |
