| # Evaluated 14B step-80 settings |
|
|
| The released configuration corresponds to the **three-encoder step-80 non-EMA model used in the human and VLM evaluations**. The earlier two-encoder / 8,852-reference configuration described a different experiment. |
|
|
| | Setting | Value | |
| |---|---| |
| | Generator | Wan2.1-T2V-14B, causal softmax attention | |
| | Initialization | Causal ODE; SHA-256 `01a21da77adf0e1e803d74ae73ac9c2bc397b1a05a74cbf89709a653282f7ab6` | |
| | Frozen encoders | DINOv3 + V-JEPA 2 + VideoMAEv2 | |
| | Loss weights / divisor | **10 : 1 : 1 / 3**, after detached per-encoder normalization | |
| | Reference videos | **8,000** | |
| | Distribution / differentiation / reference minibatch | **256 / 128 / 128** | |
| | GPUs / per-rank generated / per-rank differentiated | **8 / 32 / 16** | |
| | Physical generation / logical feature / physical replay batch | **2 / 16 / 1** | |
| | AdamW | LR `2.5e-6`, betas `(0, 0.999)`, epsilon `1e-8`, weight decay `0`, gradient clipping `10` | |
| | Schedule | Cosine to `5e-7` over **200** updates; take checkpoint **80** | |
| | Nyström / Monte Carlo mixture | **2 : 1** | |
| | Training video / causal chunk | 81 RGB frames at 480 × 832; 21 latent frames; 3 latent frames per chunk | |
| | Inference weights | `checkpoint_model_000080/model.pt`, regular `generator`, no EMA or CFG | |
|
|
| The original training config SHA-256 is `fba9c9abd26f9b6b08ed5450c97d01a250da2747c7325db469a2e2f204f694ab`. Public asset paths and descriptive identities are relocated; the three reference-pack hashes and numerical training settings are retained. The public recipe stops at update 80 without shortening the original 200-update learning-rate curve. `seed: 0` chooses an effective seed at startup; exact historical training replay additionally requires its effective seed and RNG state. |
|
|
| ## Inference |
|
|
| The pipeline and sample-batch noise order follow the previous step-80 inference run: |
|
|
| - Four trained stages `[1000, 750, 500, 250]`, warped to approximately `[1000, 937.5, 833.3333, 625]`, then a terminal `t=0` cleanup. |
| - Training-matched context-cache writes, with auxiliary noise seed `12345`. Scheduler `t=0` has a small nonzero noise level; cache input is noised separately from the saved clean output. |
| - **Seed 0, five samples per prompt**, continuous random stream. Draw all five samples' initial/transition noises together, then execute one video at a time. |
| - 81 output frames, 832 × 480, 16 FPS. |
|
|
| ```bash |
| CHECKPOINT=/path/to/checkpoint_model_000080/model.pt \ |
| PROMPTS=/path/to/prompts.txt \ |
| SEED=0 NUM_SAMPLES=5 OUTPUT_DIR=outputs/step80 \ |
| bash scripts/infer.sh |
| ``` |
|
|
| Set `NUM_SAMPLES=1` for a smaller demo. This changes the random-number sequence. Exact historical video reproduction requires the same prompts, ordering and shard assignment in addition to the same checkpoint. The original VBench source used four contiguous shards with 237 / 237 / 236 / 236 prompt rows, restarting seed 0 for each shard. This package provides generation, not human/VLM scoring infrastructure. |
|
|
| See [asset preparation](assets.md) for encoder geometry and reference bandwidths, and [validation](validation.md) for checks and limits. |
|
|
