Skip to content

Setup and System Requirements

This guide describes how to set up your environment to run the Cosmos Transfer 2.5 + FiftyOne workflow to augment the BioTrove moth dataset and explore results in the FiftyOne App.

The setup has three main parts:

  1. System and software requirements
  2. Installing Cosmos Transfer 2.5 and its dependencies
  3. Installing and configuring FiftyOne and the dataset

System Requirements

Minimum Hardware Requirements

  • GPU:
  • 1 or more NVIDIA GPUs
  • Ampere architecture or newer (e.g., RTX 30 Series, A100, H100 or later recommended)
  • Storage:
  • Sufficient disk space for:
    • Cosmos Transfer 2.5 repository and model weights
    • FiftyOne datasets and derived videos/images (edge maps, outputs, last frames)

Supported Platform

  • Operating System: Linux x86-64
  • Recommended: Ubuntu ≥ 22.04 (glibc ≥ 2.35)
  • NVIDIA Driver:
  • ≥ 570.124.06, compatible with CUDA 12.8.1 (or CUDA 12+)
  • Python:
  • Python 3.10 (aligns with Cosmos Transfer 2.5 requirements)

Software Requirements

You will need the following software components:

  • CUDA Toolkit compatible with your driver (CUDA 12+)
  • PyTorch ≥ 2.5 with CUDA support
  • TorchVision
  • Git
  • FFmpeg (CLI) – for converting images to videos and handling video I/O
  • Python packages:
  • fiftyone
  • opencv-python (for Canny edges and video manipulation)
  • Cosmos Transfer 2.5 Python package (installed from the repo)
  • Additional Cosmos dependencies (typically installed via its setup guide), such as:
    • json5
    • gradio (optional UI)
    • easyio (for multi-storage backends)

For the most accurate list of Cosmos Transfer 2.5 dependencies, always refer to the official Cosmos Transfer 2.5 Setup Guide.

Installation

1. Create and Activate a Python Environment

You can use either conda or venv. Example with conda:

conda create -n cosmos-transfer2_5-biotrove python=3.10 -y
conda activate cosmos-transfer2_5-biotrove

Or with venv:

python3.10 -m venv cosmos-transfer2_5-biotrove
source cosmos-transfer2_5-biotrove/bin/activate

2. Install FFmpeg

On Ubuntu:

sudo apt-get update
sudo apt-get install -y ffmpeg

You should be able to run:

ffmpeg -version

without errors.

3. Install FiftyOne and Supporting Python Packages

Install FiftyOne and core Python dependencies:

pip install fiftyone opencv-python

Optional, but recommended for working with notebooks and visualization:

pip install jupyterlab umap-learn

4. Clone and Install Cosmos Transfer 2.5

Clone the Cosmos Transfer 2.5 repository and install it in editable mode.

git clone https://github.com/nvidia-cosmos/cosmos-transfer2.5.git
cd cosmos-transfer2.5
pip install -e .

Then follow the Cosmos Transfer 2.5 Setup Guide for environment configuration and model weight downloads.

5. Configure Environment Variables and Paths

Set COSMOS_DIR:

export COSMOS_DIR=/path/to/cosmos-transfer2.5

Optional:

export LIST_FILE=/path/to/video_list.txt
export MAX_VIDS=100

Dataset Setup (BioTrove Moth Dataset)

This recipe uses a subset of the BioTrove dataset from the Hugging Face Hub:

dataset_src = fouh.load_from_hub(
    "pjramg/moth_biotrove",
    persistent=True,
    overwrite=True,
    max_samples=2,
)

Verification

Verify Cosmos Transfer 2.5

cd $COSMOS_DIR
python examples/inference.py --help

Verify FiftyOne + FFmpeg

import fiftyone as fo
print("FiftyOne version:", fo.__version__)

dataset_src = fouh.load_from_hub(
    "pjramg/moth_biotrove",
    persistent=True,
    overwrite=False,
    max_samples=2,
)

Minimal FFmpeg test:

ffmpeg -loop 1 -i some_image.jpg -t 1 -c:v libx264 -pix_fmt yuv420p test.mp4

Next Steps

You can go to the inference tutorial to complete the Cosmos Transfer 2.5 + FiftyOne workflow. And visit this tutorial to run it directly in your environment.