MemorizeWhenNeeded-Eval

Memorize When Needed: Decoupled Memory Control for Spatially Consistent Long-Horizon Video Generation

Yanjun Guo, Zhengqiang Zhang, Pengfei Wang, Xinyue Liang, Zhiyuan Ma, Lei Zhang

🚩 Accepted by NeurIPS 2026

Project Page | Paper | Model Weights

---

Overview

This repository provides the official evaluation data for our paper "Memorize When Needed: Decoupled Memory Control for Spatially Consistent Long-Horizon Video Generation", accepted at NeurIPS 2026.

The dataset contains 242 evaluation samples across two subsets:

• RealEstate10K (42 samples): Test samples derived from the RealEstate10K dataset, used for evaluating long-term spatial consistency and video generation quality along camera trajectories with revisiting.
• Out-of-Distribution (OOD, 200 samples): Additional test scenes covering diverse indoor and outdoor environments, sourced from internet images and AI-generated content. This subset evaluates generalization to unseen scenes and visual styles.

Both subsets provide initial images, descriptive text prompts, and camera parameters for reproducible evaluation of camera-controlled long-horizon video generation.

Dataset Structure

MemorizeWhenNeeded-Eval/
├── README.md
├── real10k/
│   ├── metadata_real10k.npz
│   ├── prompts.json
│   └── Images/
└── ood/
    ├── metadata_ood.npz
    ├── prompts.json
    └── Images/

Image filenames are preserved from the original evaluation data.

Data Format

Metadata (metadata_*.npz)

Each NPZ file contains a NumPy object array stored under the key arr_0, with one dictionary per evaluation sample.

Each dictionary contains the following fields:

Field	Description
dataset_source	Source identifier of the evaluation sample
image_path	Path to the initial image, relative to the corresponding subset directory
long_caption	Text description used as the generation prompt
camera_intrinsics	Camera intrinsic parameters, stored as a NumPy array of shape (4,)
camera_extrinsics	Camera extrinsic matrices, stored as a NumPy array of shape (200, 4, 4)

Dataset source identifiers:

• real10k_test — RealEstate10K test samples
• ood-indoor — OOD indoor scenes
• ood-outdoor — OOD outdoor scenes

Camera Parameters

Both RealEstate10K and OOD use the RealEstate10K camera convention, with extrinsics stored as World-to-Camera (W2C) transformations.

Text Prompts (prompts.json)

Each prompts.json contains an ordered list of dictionaries with three fields:

Field	Description
dataset_source	Source identifier
image_path	Relative path to the initial image
long_caption	Text prompt describing the scene

The entries appear in exactly the same order as the corresponding NPZ records, and all three fields are copied directly from the metadata.

Usage

Download

Download the complete dataset using the Hugging Face CLI:

hf download Guoyanjun/MemorizeWhenNeeded-Eval \
    --repo-type dataset \
    --local-dir ./MemorizeWhenNeeded-Eval

Load Metadata and Images

import json
from pathlib import Path
import numpy as np

root = Path("./MemorizeWhenNeeded-Eval")

# Choose "real10k" or "ood"
split = "real10k"
split_dir = root / split

# Load metadata
metadata_path = split_dir / f"metadata_{split}.npz"
with np.load(metadata_path, allow_pickle=True) as archive:
    records = archive["arr_0"]

# Load prompts
with open(split_dir / "prompts.json", "r", encoding="utf-8") as f:
    prompts = json.load(f)

# Access an evaluation sample
sample = records[0]

image_path = split_dir / sample["image_path"]
caption = sample["long_caption"]

camera_intrinsics = sample["camera_intrinsics"]
camera_extrinsics = sample["camera_extrinsics"]

assert image_path.is_file()
assert prompts[0]["image_path"] == sample["image_path"]

print("Number of samples:", len(records))
print("Image:", image_path)
print("Prompt:", caption)
print("Intrinsics shape:", camera_intrinsics.shape)  # (4,)
print("Extrinsics shape:", camera_extrinsics.shape)  # (200, 4, 4)

Note: Loading the NPZ files requires allow_pickle=True because the archives contain NumPy object arrays. Only load these files from trusted sources.

Citation

If you find this evaluation dataset useful in your research, please cite our paper:

@article{guo2026memorize,
  title={Memorize When Needed: Decoupled Memory Control for Spatially Consistent Long-Horizon Video Generation},
  author={Guo, Yanjun and Zhang, Zhengqiang and Wang, Pengfei and Liang, Xinyue and Ma, Zhiyuan and Zhang, Lei},
  journal={arXiv preprint arXiv:2604.18215},
  year={2026}
}

Acknowledgements

The RealEstate10K subset is derived from the RealEstate10K dataset. We acknowledge the original dataset authors.

Please respect the applicable licenses and usage conditions of the underlying source data.