long_rolling_plain_from_cd

OmniMem long rolling video generation with a fixed prompt, using a CD-derived self-forcing initialization.

This release contains the EMA generator checkpoint at training step 1200. The initialization lineage is CD step-4500 EMA → ordinary self-forcing step-600 EMA → plain rolling training.

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.

from omnimem.models.transformers.causal_wan_nsa_model import CausalWanNSAModel

model = CausalWanNSAModel.from_pretrained(
    "omnimem/long_rolling_plain_from_cd"
)

The original checkpoint configuration is preserved unchanged. Its saved class name records the training wrapper, and _name_or_path records the initialization source. The released weights are the step-1200 EMA weights.

Training evaluation settings

Setting	Value
Prompt mode	Fixed prompt throughout each video
Resolution	832 × 480
Video duration	60 seconds
Frame rate	16 FPS
Exported frames	960 (969 generated, tail cropped)
Sampling steps	4
Guidance scale	1
Timestep shift	5
Seed	42
Frame block size	3
KV history window	3
KV cache LRU maximum	7
VAE decode chunk size	3

Training evaluation uses five full-length static prompts.