This project is based on LTX-2 (Lightricks) and is released, as a whole, under the LTX-2 Community License Agreement (see LICENSE). It also uses code and model weights from the additional open-source projects listed below; all copyright notices are preserved in the source files, and third-party components remain under their original licenses.
Directory: flash_alaya/ltx2/
Source: Lightricks
Repository: https://github.com/Lightricks/LTX-2 · https://huggingface.co/Lightricks/LTX-2.3
License (code and model): LTX-2 Community License Agreement (dated 2026-01-05) — https://github.com/Lightricks/LTX-2/blob/main/LICENSE . This is not Apache 2.0. Free for academic use and for commercial use by entities under $10M ARR; entities with ≥ $10M ARR need a paid commercial license (ltxv-licensing@lightricks.com). Any Derivative of LTX-2 must be distributed exclusively under this same Agreement, retaining copyright notices, shipping the full license + use restrictions, and marking modified files. Includes use-based restrictions (Attachment A).
Usage: Our DiT and streaming pipeline in flash_alaya/ltx2/ are refactored/adapted from LTX-2 (derivative code). Our released weights merged_infer.safetensors are fine-tuned from the LTX-2.3-22B base, and the VAE / text-encoder tensors in that bundle are taken from the original LTX-2.3 weights — i.e. both the ltx2/ code and the released weights are derivatives of LTX-2.3 and are governed by the LTX-2 Community License Agreement, not this repo's Apache 2.0.
Directory: flash_alaya/ltx2/ (VAE and "WAN-style" module structure; see ltx2/utils/ltx2_streaming_vae.py, ltx2/configs/)
Source: Wan-Video (Alibaba)
Repository: https://github.com/Wan-Video/Wan2.1
License: Apache 2.0
Usage: The streaming VAE design and several module conventions follow Wan (the VAE is based on the Wan2.1 VAE). No GPL/strong-copyleft code is introduced.
File: flash_alaya/utils/taehv.py (vendored)
Source: madebyollin
Repository: https://github.com/madebyollin/taehv
License: MIT
Usage: Optional tiny latent decoder (--bank-taehv), taeltx2_3_wide variant matching LTX-2.3 latents. Attribution header is preserved in the source file.
Directory: checkpoints/Depth-Anything-3/ (external code repo, installed at runtime; not bundled in this repository)
Source: ByteDance-Seed
Repository: https://github.com/ByteDance-Seed/Depth-Anything-3
Source Code License: Apache 2.0
Model License: see the DA3 model card (depth-anything/DA3NESTED-GIANT-LARGE-1.1)
Usage: Provides the depth estimation used by the spatial-memory (3D cache) branch. Both the DA3 code and its weights are obtained by the user from the official sources; neither is redistributed here.
Source: Google Repository / Weights: https://huggingface.co/google/gemma-3-12b-it-qat-q4_0-unquantized License: Gemma Terms of Use (https://ai.google.dev/gemma/terms) — restricted, not an OSI-approved open-source license Usage: Used at runtime as the text encoder, loaded from a user-downloaded checkpoint. Gemma weights are not redistributed by this project; users must download them from Google/Hugging Face and accept the Gemma Terms of Use.
| Component | Source | License |
|---|---|---|
| diffusers, transformers, huggingface-hub | Hugging Face | Apache 2.0 |
| PyTorch (torch/torchvision/torchaudio) | PyTorch | BSD-3-Clause |
| xformers | Meta | BSD-3-Clause |
| NumPy, SciPy | community | BSD-3-Clause |
| OpenCV (opencv-python) | OpenCV | Apache 2.0 |
| einops | community | MIT |
| safetensors, peft | Hugging Face | Apache 2.0 |
| bitsandbytes | community | MIT |
This project is released as a whole under the LTX-2 Community License Agreement. The third-party components below are compatible with being used/included under it.
| Component | License | Note |
|---|---|---|
LTX-2 (source code — basis of ltx2/) |
LTX-2 Community License Agreement | this project's governing license |
| LTX-2.3 (model weights — base of our weights) | LTX-2 Community License Agreement | our weights are a derivative, shipped under the same license |
| Wan / Wan2.1 (source code) | Apache 2.0 | ✓ permissive |
| TAEHV | MIT | ✓ permissive |
| Depth-Anything-3 (source code) | Apache 2.0 | ✓ permissive |
| Depth-Anything-3 (model weights) | DA3 model license | not redistributed |
| Gemma-3 (model weights) | Gemma Terms of Use | not redistributed |
| diffusers / transformers / HF libs | Apache 2.0 | ✓ |
| NumPy / SciPy / PyTorch / xformers | BSD-3-Clause | ✓ |
| OpenCV | Apache 2.0 | ✓ |
| einops / bitsandbytes | MIT | ✓ |
Model weights are subject to their providers' separate license terms. Both the
flash_alaya/ltx2/ code and the released merged_infer.safetensors weights are
derivatives of LTX-2.3 and are governed by the LTX-2 Community License
Agreement. Gemma-3 and Depth-Anything-3 weights are not redistributed here.
- This project is released, as a whole, under the LTX-2 Community License Agreement (see LICENSE); it is for academic research and non-commercial use only.
- All original LTX-2 copyright, license, patent, trademark, and attribution notices are retained; modifications by Alaya Lab are documented in file headers.
- Other third-party code (Wan, TAEHV, Depth-Anything-3, framework libraries) remains under its original permissive license.
- Model weights may be subject to additional license terms from their respective providers. Gemma-3 and Depth-Anything-3 weights are not redistributed by this project.
For questions, open an issue in the project repository.