epfl-vita/PTD-weights

PTD checkpoints

Four-step and six-step MiniMax-H3 adapter checkpoints trained by EPFL VITA Lab based on Parametric Trajectory Distillation for Few-Step Video Generation (PTD), with the re-noised hop transitions of the recipes h3_4step_sde and h3_6step_sde.

Checkpoints

Directory	Recipe	Sampling steps	Training	Size
ptd_h3_4step_sde/	h3_4step_sde	4	100 optimizer steps, batch 512	11.5 GB
ptd_h3_6step_sde/	h3_6step_sde	6	100 optimizer steps, batch 512	17.2 GB

Each directory contains a torch.distributed.checkpoint state dict with per-hop LoRA adapters and training-only heads. Inference uses the adapters and ignores the heads. Training settings, hop grids and provenance are recorded in ptd_info.json; checksums/ lists the SHA-256 of every file.

Hop transitions. In the _sde recipes, the state carried from one sampling step to the next is the step's clean endpoint re-noised to the next noise level with fresh noise, instead of the Euler endpoint, in training and sampling alike (the last step is unchanged). A checkpoint is sampled with the recipe it was trained with: h3_4step_sde for ptd_h3_4step_sde/ and h3_6step_sde for ptd_h3_6step_sde/.

Previous checkpoints. This update replaces the paper's checkpoints ptd_h3_4step/ and ptd_h3_6step/ (recipes h3_4step / h3_6step, Euler hop transitions, 149 / 199 optimizer steps), which this repository held until revision 8a70e31c82.

Evaluation. Paired GPT-6 Astra judgments on 100 prompts × 3 seeds in four categories (dynamic quality, static quality, naturalness, diversity), with 95% bootstrap confidence intervals of the paired difference:

• Four-step. At equal training length (50 steps) the re-noised recipe scores +0.10 in dynamic quality over the paper's recipe (confidence interval above zero), with no other category worse. Its step-100 checkpoint, released here, equals the paper's four-step checkpoint (149 steps) in all four categories at two thirds of its training compute. Trained on to 149 steps it is not better than the paper's checkpoint and slightly softer, which is why step 100 is released.
• Six-step. At 50, 100 and 150 steps the re-noised recipe ties the paper's recipe at the same step in dynamic quality, static quality and naturalness (every confidence interval contains zero). The step-100 checkpoint is released for consistency with the four-step one; it has not been judged against the paper's six-step checkpoint (199 steps).

Usage

Code and setup instructions: PTD inference (standalone and ComfyUI).

For standalone inference, pass the checkpoint directory as --ckpt and its matching recipe from the table as --recipe. The _sde recipes re-noise the hop transitions during sampling, as the training did; they are in PTD inference from version 0.2.0 (branch release/comfyui-inference, standalone and ComfyUI paths). Older inference code with Euler transitions only (the paper's recipes h3_4step / h3_6step) loads and runs these checkpoints, but its outputs are not the ones evaluated above.

Download

Download one or both students:

hf download epfl-vita/PTD-weights --local-dir weights                                                # both students (28.6 GB)
hf download epfl-vita/PTD-weights --local-dir weights --include "ptd_h3_6step_sde/*" "checksums/*"   # only the six-step student
cd weights && sha256sum -c checksums/ptd_h3_6step_sde.files.sha256                                     # optional: check every file

The adapters were trained on MiniMaxAI/MiniMax-H3 at revision 42ed227ee7df40d41602854ae760620d6eb651fe; download the base model separately, at that revision.

License

These adapters are Model Derivatives under the MiniMax H3 Community License Agreement. See LICENSE-MiniMax-H3.txt and NOTICE for terms and attribution.

Citation

@article{feng2026ptd,
  title   = {Parametric Trajectory Distillation for Few-Step Video Generation},
  author  = {Feng, Lan and Karkus, Peter and Igl, Maximilian and Berner, Julius and Chen, Yuxiao and
             Tan, Shuhan and Alahi, Alexandre and Ivanovic, Boris and Pavone, Marco},
  journal = {arXiv preprint arXiv:2610.11498},
  year    = {2026}
}