| --- |
| license: other |
| base_model: MiniMaxAI/MiniMax-H3 |
| tags: |
| - lora |
| - video-generation |
| - world-model |
| - multi-view |
| --- |
| |
| # MVWM LoRA checkpoints |
|
|
| Research checkpoints from MVWM, a multi-person, multi-view video world model project. Every file is a LoRA (plus a few small extra modules) trained on top of the **Ref2VA** variant of [`MiniMaxAI/MiniMax-H3`](https://huggingface.co/MiniMaxAI/MiniMax-H3). The weights cannot be used on their own: you need the base model and the matching training/inference code, which is not public. The safetensors headers carry no metadata beyond `{"format": "pt"}`. |
|
|
| MVWM 多人多视角世界模型课题的 LoRA 存档（基座 MiniMax-H3 Ref2VA）。需要配合基座和对应代码使用。 |
|
|
| ## Layout |
|
|
| `<run>/step-N.safetensors` is the checkpoint after N loop steps. In the dual-DiT runs (`ckpt_gather_*`), 2 loop steps make one optimizer update. |
|
|
| | Run | Recommended step | What it is | |
| |---|---|---| |
| | `ckpt_gather_s11` | **1200** (also 800) | **Current main line.** Two views run as two sequences. In every self-attention layer, each view's keys and values also include up to K=128 tokens per frame gathered from the other view, under one softmax with no new parameters (`gather_config.json`). Not fully trained. | |
| | `ckpt_gather_s12` | 1200 | Same as s11, but with K=64. | |
| | `ckpt_gather_s14` | 400 (also 600) | s11 plus Plücker ray-map camera conditioning. Ties s11 at step 400 and is about 1 dB worse at step 600. | |
| | `ckpt_v1_kf3_cont` | **1250** (also 1812) | Stage-1 final recipe: both views stacked into one video, full attention, soft first-frame anchoring, gray-background box proxy, time-aligned proxy, rank 32. Step 1250 is the final stage-1 checkpoint (2250 steps in total). | |
| | `ckpt_v1_kf3` | — | Run that `ckpt_v1_kf3_cont` continues from. | |
| | `ckpt_v2_ha`, `ckpt_v2b_ha` | 1812 (v2b) | Stage-1 variant: hard anchoring, rank 128. Did not beat v1. | |
| | `ckpt_v3b` / `ckpt_v3b_shared` | 750 / 1002 | Stage-2 identity-colored proxy, without / with a shared-entity cross-view attention mask. | |
| | `ckpt_cwm_v3c`, `ckpt_cwm_v3c_eval` | 300 | Stage-2 CWM-style recipe; the stacked-video line stopped here. | |
| | `ckpt_v0`, `ckpt_v0_dlc` | — | Early baselines. | |
|
|
| Every file without a recommended step is an intermediate checkpoint, kept as a complete record. `ckpt_gather_s11/step-700` is missing because the original file was corrupted (0 bytes). |
|
|
| Download one run: |
|
|
| ```bash |
| hf download lee238/MVWM-weights --include "ckpt_gather_s11/*" --local-dir mvwm_weights |
| ``` |
|
|
