| --- |
| license: other |
| library_name: pytorch |
| pipeline_tag: image-to-video |
| tags: [longlive-2.0, robotics, autoregressive] |
| --- |
| |
| # LongLive 2.0 Robot-L |
|
|
| Autoregressive robot-video generation weights. |
| L continues from Robot-M with long-sequence training; this is the final Stage-L checkpoint at step 3,972. |
| S, M and L describe training sequence length, not model size. |
|
|
| ## Download |
|
|
| ```bash |
| hf download Efficient-Large-Model/LongLive2.0-Robot-L --local-dir ./weights/LongLive2.0-Robot-L |
| ``` |
|
|
| ## Inference |
|
|
| Instantiate the matching LongLive 2.0 generator, then load the FP32 weights: |
|
|
| ```python |
| import torch |
| |
| state_dict = torch.load("model.pt", map_location="cpu", mmap=True, weights_only=True) |
| model.load_state_dict(state_dict, strict=True) |
| ``` |
|
|
| Run the loading example from the downloaded directory. The matching inference |
| runtime, VAE and text encoder are required separately; this is not a standalone |
| pipeline. This release contains the online generator only (825 FP32 tensors, |
| 4,999,787,712 parameters). Optimizer, EMA and other training states are not included. |
|
|
| All 20 Stage-L fixed-video validation gates were accepted; the final panel contains |
| 24 videos evaluated with 50 sampling steps, CFG 3.0 and base seed 20260711. |
| These checks are release validation, not a robot-control safety or benchmark claim. |
| See the [Long-WAM Stage 1 guide](https://github.com/NVlabs/LongLive/tree/main/Long-WAM/stage1) |
| for the matching runtime and inference instructions. |
|
|
| Original upstream model terms continue to apply. |
|
|
