[CVPR'26 🏅Highlight] MoRel: Long-Range Flicker-Free 4D Motion Modeling via Anchor Relay-based Bidirectional Blending with Hierarchical Densification

Sangwoon Kwak^1*, Weeyoung Kwon^2*, Jun Young Jeong¹, Geonho Kim², Won-Sik Cheong¹, Jihyong Oh^2†

¹Electronics and Telecommunications Research Institute (ETRI) ²Chung-Ang University
*Equal contribution, †Corresponding author

This repository is the official PyTorch implementation of "MoRel: Long-Range Flicker-Free 4D Motion Modeling via Anchor Relay-based Bidirectional Blending with Hierarchical Densification."

---

News

• 2025.12.10: MoRel ArXiv and Project page are initially released.
• 2026.02.21: MoRel is accepted to CVPR 2026! 🔥
• 2026.03.26: MoRel ArXiv is updated (camera-ready version).
• 2026.04.09: MoRel is selected as a Highlight at CVPR 2026! ✨
• 2026.04.10: MoRel codebase is initially released.

---

Abstract

Recent advances in 4D Gaussian Splatting (4DGS) have extended the high-speed rendering capability of 3D Gaussian Splatting (3DGS) into the temporal domain, enabling real-time rendering of dynamic scenes. However, one of the major remaining challenges lies in modeling long-range motion-contained dynamic videos, where a naïve extension of existing methods leads to severe memory explosion, temporal flickering, and failure to handle appearing or disappearing occlusions over time.

To address these challenges, we propose a novel 4DGS framework characterized by an Anchor Relay-based Bidirectional Blending (ARBB) mechanism, named MoRel, which enables temporally consistent and memory-efficient modeling of long-range dynamic scenes. Our method progressively constructs locally canonical anchor spaces at key-frame time indices and models inter-frame deformations at the anchor level, enhancing temporal coherence.

By learning bidirectional deformations between Key-frame Anchors (KfA) and adaptively blending them through learnable opacity control, our approach mitigates temporal discontinuities and flickering artifacts. We further introduce a Feature-variance-guided Hierarchical Densification (FHD) scheme that effectively densifies KfAs while preserving rendering efficiency.

Method Overview

MoRel proceeds through four stages:

Stage	Name	Description
GCA	Global Canonical Anchor	Trains a single globally consistent anchor scaffold over the entire sequence
KfA	Key-frame Anchor	Periodically placed local canonical spaces, initialized from GCA and refined per key-frame
PWD	Progressive Windowed Deformation	Each KfA learns forward/backward deformation fields within a local temporal window
IFB	Intermediate Frame Blending	Learns temporal opacity weights to smoothly blend neighboring KfAs

---

Environment Setup

git clone <repository_url>
cd MoRel

# Install dependencies and create conda environment
bash install.sh

# Activate the environment
conda activate morel

The install script creates a conda environment named morel and installs all required packages including submodules (depth-diff-gaussian-rasterization, simple-knn).

---

Dataset Preparation

SelfCap_LR

Multi-camera synchronized capture dataset. Each scene contains per-frame multi-view images and per-frame sparse point clouds (from COLMAP triangulation).

Dataset: First, obtain the original SelfCap dataset from FreeTimeGS, then locate the data as described below.

data/selfcap/
└── <scene>/                       # e.g. bike, corgi, yoga, dance
    ├── images/
    │   ├── 0000/                  # camera 0000
    │   │   ├── 000000.jpg
    │   │   ├── 000001.jpg
    │   │   └── ...
    │   ├── 0001/                  # camera 0001
    │   └── ...
    ├── colmap/
    │   ├── <frame_id>/            # 6-digit frame index (e.g. 000000)
    │   │   ├── images/            # per-frame images (symlinked)
    │   │   ├── sparse/
    │   │   │   ├── cameras.txt
    │   │   │   └── images.txt
    │   │   └── masks/             # foreground masks
    │   └── ...
    └── dense_pcds/
        ├── 000000.ply             # dense point cloud at frame 0
        ├── 002000.ply
        └── ...

Note: Scenes are available in three temporal lengths — short (~300 frames), medium (~1200 frames), long (~3600 frames).

Custom Dataset

Custom multi-camera dataset. Run scripts/run_colmap_custom.sh once per scene to obtain camera parameters, then scripts/run_colmap_custom_all_frames.sh to generate per-frame sparse point clouds.

data/custom/
└── <scene>/                       # e.g. Makeup, RX0_251029_s01t01
    ├── frames/                    # (or images/)
    │   ├── cam01/
    │   │   ├── 000001.jpg
    │   │   ├── 000002.jpg
    │   │   └── ...
    │   ├── cam02/
    │   └── ...
    ├── cameras.txt                # COLMAP camera intrinsics (generated once)
    ├── images.txt                 # COLMAP camera extrinsics (generated once)
    ├── points3D.txt               # Sparse 3D points from frame 0
    ├── pcds/
    │   ├── frame_000001.ply       # Per-frame sparse point clouds
    │   ├── frame_000002.ply
    │   └── ...
    └── colmap_frame0/             # Working directory from run_colmap_custom.sh
        └── ...                    # (can be deleted after cameras.txt is generated)

COLMAP preparation steps for Custom dataset:

# Step 1: Run full SfM on the first frame to get camera parameters
bash scripts/run_colmap_custom.sh

# Step 2: Triangulate sparse point clouds for all frames
bash scripts/run_colmap_custom_all_frames.sh

Other Supported Datasets (HyperNeRF, Dycheck, DyNeRF)

HyperNeRF

Download from the HyperNeRF dataset page.

data/hypernerf/
└── <scene>/                       # e.g. 3dprinter, chickchicken
    ├── camera/
    │   ├── <frame_id>.json        # per-frame camera parameters
    │   └── ...
    ├── rgb/
    │   ├── 1x/
    │   │   ├── <frame_id>.png
    │   │   └── ...
    │   └── 2x/                    # half resolution
    ├── colmap/
    │   ├── sparse/
    │   └── dense/
    ├── dataset.json
    └── points.npy

Dycheck (iPhone)

Download from the Dycheck dataset page.

data/dycheck/
└── <scene>/                       # e.g. apple, block, spin
    ├── camera/
    │   ├── <frame_id>.json
    │   └── ...
    ├── rgb/
    │   ├── 1x/
    │   └── 2x/
    ├── depth/
    │   ├── 1x/
    │   └── 2x/
    ├── colmap/
    │   ├── sparse/
    │   └── dense/
    └── dataset.json

Note: Dycheck uses masked metrics (mPSNR, mSSIM, mLPIPS) as reported in the paper. Refer to Dycheck's evaluation modules for masked evaluation.

DyNeRF

Download from the Neural 3D Video dataset page.

data/dynerf/
└── <scene>/                       # e.g. flame_steak, coffee_martini
    ├── cam00/
    │   └── images/
    │       ├── 0000.png
    │       ├── 0001.png
    │       └── ...
    ├── cam01/
    ├── ...
    └── colmap/
        ├── sparse/
        └── dense/

---

Training & Rendering

Pre-written run scripts are located in run_scripts/. Each script interactively prompts for options.

SelfCap_LR

cd run_scripts
bash run_selfcap.sh

Example session:

Enter the GPU ID (default: 0): 0
Enter the port (default: 5000): 5000
Enter the config number (default: 1.0): 1.0

Select the scene to run:
  [bike_1]   0: short  1: medium  2: long
  [bike_2]   3: short  4: medium  5: long
  [corgi]    6: short  7: medium  8: long
  [yoga]     9: short  10: medium 11: long
  [dance]    12: short 13: medium 14: long
  [all]      short_all: 15  medium_all: 16  long_all: 17

Enter your choice (default: 0): 6
Skip training? (rendering only) [y/n, default: n]: n
Fine-tuning? [y/n, default: n]: n

This will train on corgi_short and then render train/test/interpolation views automatically.

Custom Dataset

cd run_scripts
bash run_custom.sh

Example session:

Enter the GPU ID (default: 0): 0
Enter the port (default: 5000): 5000
Enter the config number (default: 1.0): 1.0
Enter the scene name: {your_scene_name}
  (e.g. /path/to/MoRel/data/custom/{your_scene_name})
Enter the scene data path: /path/to/MoRel/data/custom/{your_scene_name}
Skip training? (rendering only) [y/n, default: n]: n
Fine-tuning? [y/n, default: n]: n

Manual Commands

You can also run each step individually:

# Training
PYTHONPATH='.' CUDA_VISIBLE_DEVICES=0 python train.py \
    -s data/selfcap/bike_short \
    --port 5000 \
    --expname "selfcap_1.0/bike_short" \
    --configs arguments/selfcap/config_1.0.py

# Rendering
PYTHONPATH='.' CUDA_VISIBLE_DEVICES=0 python render.py \
    --model_path "output/selfcap_1.0/bike_short" \
    --configs arguments/selfcap/config_1.0.py \
    --render_frame_end 300 \
    --streaming_video \
    --skip_test --skip_interpolation

# Evaluation
PYTHONPATH='.' CUDA_VISIBLE_DEVICES=0 python metrics.py \
    --model_path "output/selfcap_1.0/bike_short"

# Aggregate metrics across scenes
python collect_metric.py \
    --output_path "output/selfcap_1.0" \
    --dataset selfcap

---

Configuration

Configs are located in arguments/<dataset>/. The default configuration for all datasets is config_1.0.py.

To experiment with different hyperparameters, simply create a new config file (e.g., config_2.0.py) in the same folder and specify the version number when running the script:

Enter the config number (default: 1.0): 2.0

To override parameters for a specific scene only, create a file named <scene_name>.py in the same folder — only the parameters defined in it will overwrite the base config.

---

Results

Please visit our project page for video demos, qualitative comparisons, and quantitative results.

---

Acknowledgement

This source code is built upon prior excellent works (3DGS, 4DGS, Scaffold-GS). We sincerely appreciate the authors for their outstanding contributions.

We also sincerely thank EasyVolcap, Long Volumetric Video, and FreeTimeGS for providing the excellent raw dataset that made our long-range video experiments possible.

---

Citation

@InProceedings{kwak2026morel,
  title={MoRel: Long-Range Flicker-Free 4D Motion Modeling via Anchor Relay-based Bidirectional Blending with Hierarchical Densification},
  author={Sangwoon Kwak and Weeyoung Kwon and Jun Young Jeong and Geonho Kim and Won-Sik Cheong and Jihyong Oh},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026},
}

---

⭐ Star History