Self-Reimplemented Version of Long-LRM

Project Page

This repository contains a self-reimplemented version of Long-LRM, including the model code, as well as training and evaluation pipelines. The reimplemented version has been verified to match the performance of the original implementation.

---

Tentative TO-DO List

• Sample config files
• Script for converting raw DL3DV files into reuqired format
• Config files for training on DL3DV
• Long-LRM evaluation results on DL3DV-140 for baseline comparison
• 2D GS support
• Post-prediction optimization
• Pre-trained model weights

---

Inference script for custom data

We provide a simple script for running inference on custom data in inference.py. Remember to fill in the Long-LRM root path, the data json path, the checkpoint path and the output folder path. Note how the input cameras are normalized as in the training dataloader. Note we provide a separate pre-trained checkpoint with aspect ratio augmentation below.

---

Pre-trained model weights

Usage: put the .pt file in a folder with the same name as your config, and then put the folder into the checkpoints folder

• DL3DV 32 960x540 inputs (Table 1): download here
• DL3DV 32 inputs with aspect ratio augmentation (from 540x540 to 540x960): download here

Copyright 2025 Adobe Inc.

Model weights are licensed from Adobe Inc. under the Adobe Research License.

---

Long-LRM evaluation results

We offer the Long-LRM evaluation results on DL3DV-140 (download here), including the rendered target views, per-view metrics, and interpolated input trajectory videos, for fellow reearchers to use as a baseline. The model is re-trained with the code in this repository on DL3DV-10K (with DL3DV-140 filtered out) and achieves a mean PSNR of 24.21. Please note that except for the first inference mini-batch, subsequent inference runs complete in about 1 second.

---

Getting Started

1. Prepare Your Data

Format your dataset following the structure of the example data in data/example_data.

• Each dataset should contain one .txt file which lists the paths to the JSON files for each scene, with one path per line.

2. Configure Your Model

Create a config file in YAML format.

• Include fields for training, data, and model settings.
• You may also supply a default config file to main.py, fields of which will be overwritten by the conflicting values in the custom config file. This is handy for running multiple experiments with only a few config changes.

3. Train or Evaluate the Model

Run sh create_env.sh to install the required packages. Use torchrun to launch the training loop:

torchrun --nproc_per_node $NUM_NODE --nnodes 1 \
         --rdzv_id $JOB_ID --rdzv_backend c10d --rdzv_endpoint localhost:$PORT \
         main.py --config path_to_your_config.yaml \
         --default-config path_to_your_default_config.yaml

Switch to Evaluation Mode

To run the evaluation loop, add the --evaluation flag to the command line:

torchrun --nproc_per_node $NUM_NODE --nnodes 1 \
         --rdzv_id $JOB_ID --rdzv_backend c10d --rdzv_endpoint localhost:$PORT \
         main.py --config path_to_your_config.yaml \
         --default-config path_to_your_default_config.yaml \
         --evaluation

---