Add README for human2robot_finetune
Browse files
human2robot_finetune/README.md
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Human2Robot Finetune (Final)
|
| 2 |
+
|
| 3 |
+
This is the **final version** of the human-to-robot video generation model, finetuned for multi-view robot manipulation output.
|
| 4 |
+
|
| 5 |
+
## Model Description
|
| 6 |
+
|
| 7 |
+
- **Input**: Human hand manipulation video + text prompt (3 prompt styles supported)
|
| 8 |
+
- **Output**: 3-view robot manipulation video
|
| 9 |
+
- **Base model**: Wan2.2-TI2V-5B
|
| 10 |
+
|
| 11 |
+
## Best Checkpoint
|
| 12 |
+
|
| 13 |
+
> **`step=21100.ckpt`** is the best performing checkpoint.
|
| 14 |
+
|
| 15 |
+
## Checkpoints
|
| 16 |
+
|
| 17 |
+
| Checkpoint | Steps | Note |
|
| 18 |
+
|---|---|---|
|
| 19 |
+
| step=20000.ckpt | 20000 | |
|
| 20 |
+
| step=20100.ckpt | 20100 | |
|
| 21 |
+
| ... | ... | |
|
| 22 |
+
| **step=21100.ckpt** | **21100** | **Best** |
|
| 23 |
+
| step=21200.ckpt | 21200 | |
|
| 24 |
+
| step=21300.ckpt | 21300 | |
|
| 25 |
+
|
| 26 |
+
## Directory Structure
|
| 27 |
+
|
| 28 |
+
```
|
| 29 |
+
human2robot_finetune/
|
| 30 |
+
├── checkpoints/ # Model checkpoints (step=20000 ~ step=21300)
|
| 31 |
+
├── val_samples/ # Validation sample videos
|
| 32 |
+
│ └── keyframe_comparison/ # Keyframe comparison images
|
| 33 |
+
└── README.md
|
| 34 |
+
```
|