| 1 | # 4-Step WM Test Cases |
| 2 | |
| 3 | Each case is a 4-step first-frame I2V test from WBench Navi. `case.json` |
| 4 | contains only public semantic controls; model paths and generated media stay |
| 5 | outside this directory. |
| 6 | |
| 7 | The case directory names match their WBench IDs. Each contains an `input.jpg` |
| 8 | and a compact `case.json` with the original WBench caption in `prompt`, plus |
| 9 | `action` and optional `fov_deg`/`seed`. The caption is passed directly to the |
| 10 | Gemma text encoder; no runtime prompt enhancement or six-field rewrite is |
| 11 | applied. |
| 12 | |
| 13 | From the `echo_wm` directory, run one case with: |
| 14 | |
| 15 | ```bash |
| 16 | python scripts/run_wm_case_causal.py \ |
| 17 | --case examples/wm_causal_cases/0079 \ |
| 18 | --checkpoint checkpoints/echo-wm-flash.safetensors \ |
| 19 | --gemma-path checkpoints/gemma-3 |
| 20 | ``` |
| 21 | |
| 22 | Use `--list` to see available cases. Use `--dry-run` to print the resolved |
| 23 | `inference_wm_causal.py` command without loading a model. |
| 24 | |
| 25 | Each checked-in case has four 96-frame actions: 4 seconds per action at 24 fps, |
| 26 | or 385 output frames including the first frame. |
| 27 | |
| 28 | | Case | Motion | Input | |
| 29 | | --- | --- | --- | |
| 30 | | `0024` | yaw right/forward alternating | `input.jpg` | |
| 31 | | `0075` | forward four times | `input.jpg` | |
| 32 | | `0079` | yaw right four times | `input.jpg` | |
| 33 | | `0081` | pitch down/up, then backward/forward | `input.jpg` | |
| 34 | | `0122` | strafe left/right alternating | `input.jpg` | |
| 35 | | `0170` | forward/backward, strafe left, yaw right | `input.jpg` | |
| 36 | |
| 37 | The shared WBench camera and bounded-cache defaults live in |
| 38 | `configs/inference_wm_causal.yaml`. |
| 39 | |
| 40 | To run a fast validation instead of the default 16-second case: |
| 41 | |
| 42 | ```bash |
| 43 | python scripts/run_wm_case_causal.py \ |
| 44 | --case examples/wm_causal_cases/0079 \ |
| 45 | --checkpoint checkpoints/echo-wm-flash.safetensors \ |
| 46 | --gemma-path checkpoints/gemma-3 \ |
| 47 | --num-frames 25 --width 256 --height 128 |
| 48 | ``` |
| 49 |