sanity2025/FadeMem-FT

FadeMem-FT Model Weights

FadeMem-FT is the released LoRA checkpoint for FadeMem, a distance-aware KV memory consolidation method for autoregressive long-video diffusion.

This repository contains the fine-tuned LoRA checkpoint only. It is not a standalone text-to-video model. Inference also requires the LongLive-1.3B base checkpoint and the Wan2.1-T2V-1.3B model files.

Files

• model.pt: FadeMem training checkpoint at step 4,000. The FadeMem inference loader reads the generator_lora state dictionary from this file.
• inference.yaml: inference configuration compatible with the official FadeMem code release.

Installation

Clone the official implementation and install its dependencies:

git clone https://github.com/aniki-ly/FadeMem.git
cd FadeMem

conda create -n fademem python=3.10 -y
conda activate fademem

pip install torch==2.8.0 torchvision==0.23.0 \
  --index-url https://download.pytorch.org/whl/cu128
pip install -r requirements.txt
pip install flash-attn==2.8.3 --no-build-isolation

Adjust the PyTorch installation command if your CUDA version differs.

Download Weights

From the FadeMem repository root, download the Wan2.1 and LongLive base files:

bash scripts/download_models.sh inference

Then download this repository into fademem_models/FadeMem-FT:

hf download sanity2025/FadeMem-FT \
  --local-dir fademem_models/FadeMem-FT

The relevant files should have the following layout:

FadeMem/
|-- fademem_models/FadeMem-FT/
|   |-- inference.yaml
|   `-- model.pt
|-- longlive_models/models/longlive_base.pt
`-- wan_models/Wan2.1-T2V-1.3B/

Inference

Add one text prompt per line to prompts/example.txt, then run:

bash scripts/infer.sh fademem_models/FadeMem-FT/inference.yaml

The provided configuration keeps the LongLive base checkpoint unchanged and loads FadeMem-FT through:

generator_ckpt: longlive_models/models/longlive_base.pt
lora_ckpt: fademem_models/FadeMem-FT/model.pt

For prompt-level multi-GPU inference:

NUM_GPUS=2 bash scripts/infer.sh fademem_models/FadeMem-FT/inference.yaml

Use at least one prompt per GPU and make the prompt count divisible by NUM_GPUS. Generated videos are written to outputs/inference/.

Checkpoint Details

• Backbone: Wan2.1-T2V-1.3B in the LongLive v1.0 model format
• Adapter: LoRA, rank 256, alpha 256
• Training step: 4,000
• Resolution: 480 x 832
• Latent block size: 3 frames
• FadeMem summary slots: 12
• Temporal allocation: power law, beta 0.3
• Local attention window: 3 frames

The checkpoint also retains the training-time critic_lora state dictionary. The official inference loader ignores it and loads only generator_lora.

License

The FadeMem code is released under Apache-2.0. This checkpoint is derived from LongLive-1.3B and Wan2.1-T2V-1.3B; users must also follow the licenses and usage terms of those upstream models. In particular, LongLive-1.3B is released under a non-commercial Creative Commons license on its Hugging Face repository.

Citation

@article{lu2026fademem,
  title   = {FadeMem: Distance-Aware Memory Consolidation for Autoregressive Video Diffusion},
  author  = {Lu, Yu and Yang, Junjie and Koniusz, Piotr and Song, YuXin and Yang, Yi},
  journal = {arXiv preprint arXiv:2606.10671},
  year    = {2026}
}