Install the HFTrainer repository before running the commands below. This repository hosts the processed pretrained base; the public-data training outputs are separate. Artifact provenance.
Wan 2.1 · T2V 1.3B
Text-to-video generation with repository-owned model, trainer and inference code.
Verified: 20 real-data training steps, saved checkpoint and native checkpoint-only base inference. Convergence is not established.
All models · Settings · Train · Infer · Evidence · Demos
At a glance
| Property | Released setting |
|---|---|
| Model | Wan 2.1 T2V · 1.3B |
| Training | Full transformer; frozen UMT5 and VAE |
| Training input | 128 × 128 × 17 frames; video only |
| Public dataset | crema_d_mini |
| Runtime | Local HFTrainer implementation; supporting PyTorch/media libraries and model assets remain dependencies |
Sources
Original paper / report · Original code
The original repository is provenance, not a runtime checkout requirement. Third-party notices preserve implementation and asset terms.
Settings and checkpoints
| Setting | Training config | Processed checkpoint | Access |
|---|---|---|---|
| Wan 2.1 T2V · 1.3B | config.py | HFTrainer-Wan2.1-T2V-1.3B | Public |
Only the setting above is a released HFTrainer artifact. Its root config.json, weights and all task-required component configs, processors and schedulers are included together. Inference needs only the checkpoint path or Hub ID, plus task inputs.
Pinned weight revision: f4a5bc6cd9b9. Original conversion evidence describes the initial export; the current revision adds checkpoint-only dispatch metadata without changing tensors.
Setup
Run from the repository root:
python -m pip install -e .
python -m pip install "huggingface_hub>=0.34,<2" "av>=14,<17"
Verified on one NVIDIA H200 with PyTorch 2.8.0+cu128. This is the tested environment, not a measured minimum-memory requirement. Install system FFmpeg (ffmpeg and ffprobe) for media preparation.
Data
The recipe downloads ZeyuLing/hftrainer_crema_d_mini to data/hftrainer_crema_d_mini. Split counts: 24 train / 8 validation / 0 test. Dataset source, license, transformations and checksum verification are documented in Public demo datasets.
Model features are cached locally when required. The script never substitutes synthetic samples for missing real media.
Train
One command downloads pinned data and weights, prepares required caches, and runs the 20-step recipe:
python tools/run_public_demo.py wan
Reuse downloaded weights with --checkpoint path/to/complete_bundle; change the output directory with --work-dir path/to/run. The underlying command is python tools/train.py configs/public_data/wan.py. Seed: 42. Training logs and checkpoints are saved under work_dirs/public_data/wan/. If you reused a custom checkpoint directory, set HFTRAINER_CHECKPOINT to that directory before directly invoking training, resume or export commands.
The final resumable checkpoint is work_dirs/public_data/wan/checkpoint-iter_20. For a longer resumed run, keep the same checkpoint/data setting:
python tools/train.py configs/public_data/wan.py --auto-resume \
--cfg-options train_cfg.max_iters=40
Infer
Run the processed pretrained base, independently of any training config:
python tools/infer.py --model ZeyuLing/HFTrainer-Wan2.1-T2V-1.3B \
--revision f4a5bc6cd9b9bdc801eac0585963b0365ab22e9f --device cuda --seed 42 \
--prompt "A butterfly rests on a flower. The camera is steady in natural daylight." \
--height 480 --width 832 --num-frames 17 --num-steps 50 --output outputs/wan.mp4
--model also accepts a local complete checkpoint directory. A resumable training checkpoint is not a standalone model: export with tools/export_model.py using the matching training config, then pass the exported directory to --model. Artifact and configuration contract.
To package this training run as a complete checkpoint (including its frozen base components):
python tools/export_model.py --config configs/public_data/wan.py \
--checkpoint work_dirs/public_data/wan/checkpoint-iter_20 --output exports/wan
Then run the inference command above with --model exports/wan and omit --revision. Complete exports duplicate the base weights; allow sufficient disk space.
Evidence and loss
Complete 1.3B transformer, UMT5 and causal VAE weights load strictly. Full components match across export/reload; tiny UMT5/VAE reference outputs match exactly, and tiny transformer FP32 maximum error is 3.58e-7.
Raw training record · Training log
These are raw, unsmoothed objectives on the public dataset. Twenty steps establish pipeline execution, not convergence. Loss values are not comparable across models.
No held-out perceptual quality metric was measured for this short run.
Demos
Play/download the generated video. Untouched published base, seed 42; prompt and sampler settings match the inference command above. Video-only preview: 480 × 832, 17 frames; CFG 5 and flow shift 3. This is separate from the short public-data fine-tuning run.
Limitations
The public training run is reduced-resolution and uses only CREMA-D video, not its audio. The preview uses 480 × 832, 17 frames, 50 Euler steps, CFG 5 and flow shift 3. Euler sampling is not the official UniPC/DPM++ solver, so no official quality-parity result is claimed. UMT5 tokenization uses an NFKC approximation of the precompiled SentencePiece normalizer; arbitrary-text parity is unestablished. The 14B and image-to-video variants are not packaged.
Citation
@misc{wan2025,
title = {Wan: Open and Advanced Large-Scale Video Generative Models},
author = {Team Wan and others},
year = {2025},
eprint = {2503.20314},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2503.20314}
}
Also retain the dataset citation and attribution.