Jiaxing Li1,2*,
Kai Zou1*,
Cindy Zhou1,3,
Kaichen Huang1,2,
Junyao Gao1,
Zile Wang1,
Yang Liu1,
Bin Liu1,
Bo An2,
Yangguang Li1†
1Riemann Dynamics 2Nanyang Technological University 3Wellington College, UK
*Equal contribution †Corresponding author
DistillAlign aligns and balances the mode-covering and mode-seeking objectives of the multi-stage autoregressive video distillation pipeline — using only a 1.3B DMD teacher, it already surpasses baselines refined with a 14B DMD teacher.
Columns · Wan2.1-1.3B | Self-Forcing | Causal-Forcing | Ours (1.3B teacher) | Ours (14B teacher)
- [2026-07] Initial release: inference code, released checkpoints, training data, and the full teacher-normalized distribution-evaluation pipeline.
All released generators are Wan2.1-1.3B students; the size in each name refers to the teacher used during training.
Checkpoints
| Model | Checkpoint | Description |
|---|---|---|
| Initializer (1.3B teacher) | 🤗 Huggingface | Pre-DMD initializer, trained toward a Wan2.1-T2V-1.3B teacher |
| Initializer (14B teacher) | 🤗 Huggingface | Pre-DMD initializer, trained toward a Wan2.1-T2V-14B teacher |
| Distilled (1.3B teacher) | 🤗 Huggingface | Final joint-distilled generator, Wan2.1-T2V-1.3B DMD teacher |
| Distilled (14B teacher) | 🤗 Huggingface | Final joint-distilled generator, Wan2.1-T2V-14B DMD teacher |
Distillation data (prompts curated from VidProM, 25K samples each):
| Data | Download | Description |
|---|---|---|
| Wan2.1-14B data | 🤗 Huggingface | 25K distillation samples generated by Wan2.1-T2V-14B |
| Wan2.1-1.3B data | 🤗 Huggingface | 25K distillation samples generated by Wan2.1-T2V-1.3B |
Other intermediate results (teacher-normalized distribution evaluation):
| Cache | Download | Description |
|---|---|---|
| Wan2.1-1.3B teacher reference features | 🤗 Huggingface | 256 x 2560 V-JEPA2 features, ready for --teacher-features |
| Wan2.1-14B teacher reference features | 🤗 Huggingface | 256 x 2560 V-JEPA2 features, ready for --teacher-features |
We recommend Linux, Python 3.10, CUDA, and:
- at least 24 GB GPU memory for 1.3B student inference;
- an 80 GB GPU for the native Wan2.1-14B teacher stage;
- at least 128 GB system memory for loading the 14B teacher shards.
Create the environment:
conda create -n distillalign python=3.10 -y
conda activate distillalign
pip install torch==2.5.1 torchvision==0.20.1 \
--index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt
pip install flash-attn --no-build-isolation
pip install -e .Download the Wan2.1 student/teacher models and the V-JEPA2 encoder:
bash scripts/download_models.shThe script writes model weights under wan_models/ and Hugging Face artifacts
under .cache/huggingface/. Both directories are excluded from Git.
Generate a video from a text prompt:
python inference.py \
--config_path configs/self_forcing_dmd.yaml \
--checkpoint_path checkpoints/distillalign_distill_14b_teacher.pt \
--prompt "A young couple stands side by side under a vast, starry night sky, gazing upward in wonder as distant stars flicker with soft, ethereal light. Their faces glow with awe, hands lightly touching, shoulders relaxed. The camera slowly pans around them, tilting upward to capture the infinite cosmos, then gently zooms in on their eyes reflecting the stars. Shot in cinematic 4K HD, with deep blues and warm golden hues, the scene pulses with quiet romance and cosmic wonder as gentle wind rustles their hair" \
--output_path outputs/inference_case.mp4For batch generation, pass --data_path prompts.txt (one prompt per line)
and --output_folder videos/ instead of --prompt/--output_path.
The evaluation compares a generator against a target teacher (Wan2.1-T2V-1.3B or Wan2.1-T2V-14B) on a fixed prompt/seed grid and reports PRDC precision and coverage in V-JEPA2 feature space.
Build the shared prompt/seed eval set (paper protocol: 16 prompts x 16 seeds,
built-in defaults, written to outputs/eval_jobs.jsonl):
python evaluate_distribution.py make-jobsFor a quick one-sample pipeline check before a full run, build a
single-sample job list instead:
python evaluate_distribution.py make-jobs --num-prompts 1 --seeds 11 --output outputs/smoke_jobs.jsonl.
Sample the evaluated checkpoint, optionally refine (initializer protocol: re-noise each endpoint to 90% and let the teacher finish denoising; skip for post-DMD checkpoints), then encode with V-JEPA2:
python evaluate_distribution.py sample \
--checkpoint checkpoints/distillalign_distill_14b_teacher.pt \
--jobs outputs/eval_jobs.jsonl \
--output-dir outputs/my_eval/student_raw
# optional refinement (initializer protocol)
python evaluate_distribution.py teacher-refine \
--raw-manifest outputs/my_eval/student_raw/manifest.jsonl \
--output-dir outputs/my_eval/student_refined \
--teacher-model Wan2.1-T2V-14B
# encode: use student_refined when refined, student_raw otherwise
python evaluate_distribution.py extract \
--manifest outputs/my_eval/student_refined/manifest.jsonl \
--output outputs/my_eval/student_refined/student_refined_vjepa2.npz \
--cache-dir .cache/huggingfaceThe checkpoint can also be a Hub URL such as
hf://LiJiaxing/DistillAlign/checkpoints/distillalign_distill_14b_teacher.pt.
Matched 25-step teacher sampling on the same prompt/seed grid, then encode:
python evaluate_distribution.py teacher-sample \
--raw-manifest outputs/my_eval/student_raw/manifest.jsonl \
--output-dir outputs/my_eval/teacher_reference \
--teacher-model Wan2.1-T2V-14B
python evaluate_distribution.py extract \
--manifest outputs/my_eval/teacher_reference/manifest.jsonl \
--output outputs/my_eval/teacher_reference/teacher_reference_vjepa2.npz \
--cache-dir .cache/huggingfaceWe release ready-made teacher feature caches for Wan2.1-T2V-1.3B and Wan2.1-T2V-14B on the paper prompt/seed grid, which replace this step entirely.
PRDC precision and coverage with k=5:
python evaluate_distribution.py metrics \
--teacher-features outputs/my_eval/teacher_reference/teacher_reference_vjepa2.npz \
--student-features outputs/my_eval/student_refined/student_refined_vjepa2.npz \
--output outputs/my_eval/precision_coverage.jsonThe metric CLI rejects feature caches whose V-JEPA2 protocol or prompt/seed set do not match.
run chains all of the above with the initializer protocol defaults:
python evaluate_distribution.py run \
--checkpoint checkpoints/distillalign_distill_14b_teacher.pt \
--jobs outputs/eval_jobs.jsonl \
--teacher-model Wan2.1-T2V-14B \
--output-dir outputs/my_evalUseful variants: --no-refine (post-DMD protocol) and --teacher-features PATH (reuse a released teacher cache).
This implementation uses Self-Forcing and Causal-Forcing as its code base, with the V-JEPA2 encoder powering the distribution evaluation. We thank the open-source community for the infrastructure that made this release possible.
If you find DistillAlign useful, please consider citing:
@misc{li2026distillalign,
title = {DistillAlign: Coordinating Mode Covering and Mode Seeking in Autoregressive Video Distillation},
author = {Li, Jiaxing and Zou, Kai and Zhou, Cindy and Huang, Kaichen and Gao, Junyao and Wang, Zile and Liu, Yang and Liu, Bin and An, Bo and Li, Yangguang},
year = {2026},
eprint = {2607.26811},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2607.26811}
}For questions and discussion, please open a GitHub issue or contact:
- Jiaxing Li — lijiaxing0213@gmail.com
- Kai Zou — kzou@mail.ustc.edu.cn
The code is released under the Apache License 2.0. Third-party model weights remain subject to their own licenses.