LTX-2 Code v1.6.0, Code Moves to Apache 2.0… Weights Keep a Separate License
Posted As of
Lightricks updated the LTX-2 repository to v1.6.0 (November 4) and v1.6.1 (November 5). The code is now Apache 2.0, but model weights are not covered by that license, and several legacy APIs were removed.
What happened
Based on Lightricks' LTX-2 GitHub release notes. This is not a new model weights release but a code-side change.
v1.6.0 (November 4)
- The repository code is now licensed under Apache 2.0. Model weights are not included in this license. LTX-2.5 keeps the LTX-2.x Community License, and LTX-2.3 keeps the LTX-2 Community License. (Official)
- Removed APIs: the context arguments of the denoiser constructor (
v_context,a_context), the resolution and frame arguments of the initial latent helper (it now takespixel_shape), the legacy model builder (replaced bySimpleBuilderandStreamingBuilder), the trainer'stext_to_videoandvideo_to_videostrategies (merged intoflexible), the trainer's automatic Hugging Face upload, and resuming training from old-format checkpoints. (Official) - Added: ProRes 4444 output, trainer strategy registration, and a unified preprocessing script. (Official)
v1.6.1 (November 5)
- Added ACES2065-1 (AP0) input and output for HDR and EXR workflows, and fixed a negative clamping error in ACEScct conversion. (Official)
DubItPipelineandTI2VidTwoStagesHQPipelineno longer enable inference mode internally. Callers in Python must wrap them withtorch.inference_mode(). (Official)
Not in the release notes: any figures for speed or VRAM changes.
Our take
If you call LTX-2 pipelines directly in code, you need to review your code before updating. The code being Apache 2.0 does not mean the weights can be used commercially. You must read the separate community license terms for the weights.
Sources and related posts
- LTX-2 GitHub 릴리스 노트 (v1.6.0, v1.6.1) Primary source