---
license: apache-2.0
base_model: Wan-AI/Wan2.2-TI2V-5B-Diffusers
tags: [video-prediction, world-model, robocasa, wan2.2, dit]
---

# WAM DiT4DiT — RoboCasa Kitchen / Wan2.2 — `5latin2latout`

Video-only (world-model) finetune of **Wan2.2-TI2V-5B** on RoboCasa Kitchen.

## Configuration
| | |
|---|---|
| training_mode | `video` (no action head — action DiT is **not** trained) |
| latent conditioning | `5latin2latout` (cond latent frames -> 2 future latent frames) |
| video EMA | saved (`_video_ema_model.*`, 825 tensors, fp32) |
| EMA schedule | diffusers `use_ema_warmup` (inv_gamma=1.0, power=0.75, cap 0.9999) |
| extraction | `WAM_VIDEO_EMA_TRAIN_EXTRACT=0` — EMA is tracked+saved only, training math unchanged |
| GPUs | 8 x H200 |

## Contents
Milestone checkpoints are stored per step under `checkpoint-<step>/`.

**Optimizer state (`global_step*`) is intentionally excluded**, so these checkpoints are
for **inference / feature extraction / probing only** — they cannot resume training.

Each folder contains the model shards (`model-*.safetensors` + index) and processor/config
files. The EMA copy of the video DiT is included in the same shards under the
`_video_ema_model.*` prefix.

## Base model & license
Derived from [Wan-AI/Wan2.2-TI2V-5B-Diffusers](https://huggingface.co/Wan-AI/Wan2.2-TI2V-5B-Diffusers);
the base model's license terms apply to this derivative.