README.md

---
pipeline_tag: text-to-video
tags:
  - omnimem
  - wan
  - video-generation
  - self-forcing
---

# self_forcing_from_ode_init

OmniMem ordinary self-forcing checkpoint initialized from ODE-trained generator weights.

This release contains the **generator EMA weights at self-forcing training step 800**.
It is the shared initialization checkpoint used for the plain and prompt-switching rolling-training branches.

## Checkpoint lineage

ODE step-800 generator weights → ordinary self-forcing step-800 EMA → rolling-training branches.

The downstream releases are:

- [long_rolling_plain_from_ode_init](https://huggingface.co/omnimem/long_rolling_plain_from_ode_init)
- [long_rolling_switch_prompt_from_ode_init](https://huggingface.co/omnimem/long_rolling_switch_prompt_from_ode_init)

## Contents

- `config.json`
- Two `safetensors` weight shards and their index
- `SHA256SUMS` for payload integrity checking

The weights retain their original FP32 storage. The model configuration uses NSA top-12 block selection.

## Usage

Use compatible OmniMem code with `CausalWanNSAModel` and `CausalWanT2VPipeline`.
The Wan2.1 VAE, UMT5 text encoder, and tokenizer are separate required components.

```python
from omnimem.models.transformers.causal_wan_nsa_model import CausalWanNSAModel

model = CausalWanNSAModel.from_pretrained(
    "omnimem/self_forcing_from_ode_init"
)
```

The original checkpoint configuration is preserved unchanged. Its saved class name
records the training wrapper, and `_name_or_path` records the upstream ODE initialization source.
The released payload is the **ordinary self-forcing step-800 EMA** checkpoint.