| --- |
| library_name: peft |
| tags: |
| - lora |
| - video-generation |
| - motion-control |
| --- |
| |
| # 4D-R1 |
|
|
| LoRA adapters for 4D-R1, a reinforcement post-training framework for motion control and efficient video generation built on VerseCrafter. |
|
|
| | Directory | Paper model | |
| |---|---| |
| | `4D-R1-GRPO` | GRPO | |
| | `4D-R1-NFT` | NFT | |
| | `4D-R1-GRPO-MP` | GRPO (mixed-precision) | |
| | `4D-R1-NFT-MP` | NFT (mixed-precision) | |
|
|
| Each directory contains `adapter_model.safetensors` and `adapter_config.json`. The adapters use LoRA rank 16 on the q, k, v, and o projections. |
|
|
| Inference requires the VerseCrafter base model, its companion components, and the compatible 4D-R1 inference implementation. |
|
|
| MP denotes mixed-precision training. |
|
|
| The corresponding paper rows are GRPO, NFT, GRPO (mixed-precision), and NFT (mixed-precision). |
|
|
| `SHA256SUMS` provides checksums for the adapter weights and configuration files. |
|
|
