You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This research checkpoint is available only after manual review by the repository owner. Describe your affiliation and intended use accurately. Approval is not automatic.

Log in or Sign Up to review the conditions and access this model content.

LTX2.5 One-Step Refiner - Step 835 EMA

This repository contains the consolidated BF16 EMA generator checkpoint produced by our LTX2.5 one-step DMD+projected-DiT-GAN+IDA training run.

Access to the files is gated and requires manual approval by the repository owner. Submitting a request does not grant access automatically.

Included artifact

File Description SHA-256
step_00000835_ema_bf16.safetensors Video-only step-835 EMA generator, BF16 2a6fff7945dae9bea146525a0eaf07d76dbb323904c32ae4f484aac29142be88
rl_lora/step_000400.pt Upstream ReFL/RL step-400 LoRA required by the current-best composition 6db8a5045e2bf7cd445fa7e0547be02e283bdd4c5fbe12e3e52c53155d1fa725

The ReFL/RL LoRA was produced by the upstream shuchenx release and is included here only because it is required by the documented current-best inference composition. We do not claim authorship of its training.

This repository does not include:

  • official LTX2.5 VAE, text encoder, base-model, or distilled-LoRA weights;
  • the upstream ReFL training source.

Training summary

  • Initialization: LTX2.5 22B DEV RC3 plus official distilled refiner LoRA-450 at 0.8.
  • Training objective: DMD + projected DiT GAN + IDA.
  • Training sigma: [0.725, 0.0].
  • GAN weight: 0.05.
  • aR1 weight: 50.
  • EMA decay: 0.98.
  • Training topology: 2 nodes x 8 H100 GPUs.
  • Selected checkpoint: step 835 EMA, chosen by visual evaluation rather than final-step selection.

Current-best composition

The current-best inference configuration described by the project is:

this step-835 EMA checkpoint
  + included upstream ReFL step-400 LoRA
  + one-step sigma [0.9093750119, 0.0]
  + official LTX2.5 video VAE / learned latent x2 upsampler

Do not fuse the official distilled LoRA again at inference time. It is already contained in the consolidated step-835 EMA checkpoint.

The checkpoint is video-only. For external MP4 refinement, retain or remux the source audio rather than describing this model as an audio refiner.

Documentation and code

The full training and inference contract is documented here:

https://github.com/Efficient-Large-Model/Sana/blob/dev/refiner_training/dev/tian/docs/streaming_causal_refiner/df_refiner/LTX25_ONE_STEP_REFINER_CURRENT_BEST.md

Training and inference code is maintained in the same Sana branch. Users must obtain all official dependencies and any unbundled third-party components from their respective upstream releases and comply with all applicable licenses and terms.

Scope and limitations

This model is optimized for perceptual detail enhancement, not conservative pixel-level restoration. It may change local geometry, motion phase, subject scale, or composition. Evaluate it on representative content before deployment.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support