Vividha: The Four Forms of Intelligence

Summary

Vividha (Sanskrit for Diversity or Multiformity) is a multimodal dataset that unifies Text, Image, Audio, and 3D modalities with quad-modality alignment for every sample, ensuring that each record contains semantically consistent representations of the same concept.

This dataset is curated by using 3D assets from Objaverse as anchors and aligning them with semantically corresponding images and audio clips from various sources using a high-precision embedding search pipeline (FAISS + EmbeddingGemma).

---

Tasks

• Audio-to-3D: Generating or retrieving 3D assets based on acoustic cues.
• Text-to-Everything: A unified prompt can simultaneously ground a concept in visual, acoustic, and spatial domains.
• Image-to-Audio: Understanding the potential sound profile of a static scene.
• Cross-Modal Capabilities: Generating and Understanding any-to-any modalities.

---

Dataset Creation

Current datasets typically align pairs of modalities (Image-Text, Audio-Text, or 3D-Text). Vividha provides a unified resource where a single concept is grounded in all four primary sensory modalities. This enables research into "Any-to-Any" generation and reasoning.

Alignment Pipeline

1. Anchor Selection: 3D objects serve as the ground truth anchors.
2. Semantic Embedding: The caption of each 3D object is encoded using google/embeddinggemma-300m.
3. Cross-Modal Search: We search pre-built FAISS indices of Image and Audio descriptions to find the semantically nearest neighbors.
4. Thresholding: Matches are kept only if similarity > 0.7.
5. Unification: The binary data for all valid matches is aggregated into a single Parquet row.

---

Dataset Structure

Each row in the dataset represents a single semantic concept (e.g., "A vintage steam locomotive") manifesting in four modalities.

Data Fields

{
    "id": "string",                 // Sequential 9-Digit ID
    "objaverse_uid": "string",      // Original Objaverse UID
    "text": "string",               // Sequential Image-Audio-3D caption
    "image": {
        "bytes": "binary",          // Binary IMAGE data
        "path": "string"            // Filename hint (e.g., "id.jpg")
    },
    "audio": {
        "bytes": "binary",          // Binary AUDIO data
        "path": "string"            // Filename hint (e.g., "id.wav")
    },
    "mesh": {
        "bytes": "binary",          // Binary GLB data
        "path": "string"            // Filename hint (e.g., "uid.glb")
    },
    "metadata": {
        "image_score": "double",
        "audio_score": "double",
        "tags": [
            "string"
        ]
    }
}

Example

{
    "id": "000000150",
    "objaverse_uid": "6afb4d9a48cf400ba1964d700e717261",
    "text": "A green train traveling along down  the tracks. A train moving. A green train.",
    "image": {
        "bytes": "b'\\xff\\xd8\\xff\\xe0\\x00\\x10JFIF\\x00\\x01\\x01\\x01\\x01,\\x01,\\x00... (truncated)",
        "path": "000000150.jpg"
    },
    "audio": {
        "bytes": "b'fLaC\\x00\\x00\\x00\"\\t\\x00\\t\\x00\\x00\\x0b\\xc3\\x00\\x10\\xda\\x07\\... (truncated)",
        "path": "000000150.flac"
    },
    "mesh": {
        "bytes": "b'glTF\\x02\\x00\\x00\\x00\\xdch\\x15\\x00D\\x83\\x01\\x00JSON{\"access... (truncated)",
        "path": "6afb4d9a48cf400ba1964d700e717261.glb"
    },
    "metadata": {
        "image_score": 0.8840261697769165,
        "audio_score": 0.7966737747192383,
        "tags": []
    }
}

---

Usage

pip install datasets trimesh soundfile pillow

1. Loading the Dataset

from datasets import load_dataset

dataset = load_dataset(
    "VINAYU7/Vividha",
    split="train",
)

sample = next(iter(dataset))

print(f"Caption: {sample['text']}")
print(f"Image size: {len(sample['image']['bytes'])} bytes")
print(f"Audio size: {len(sample['audio']['bytes'])} bytes")
print(f"Mesh size: {len(sample['mesh']['bytes'])} bytes")

2. Decode Data

from PIL import Image
import io
import trimesh
import soundfile as sf

# Decode image
image = Image.open(io.BytesIO(sample['image']['bytes']))

# Decode mesh
mesh = trimesh.load(io.BytesIO(sample['mesh']['bytes']), file_type='glb')

# Audio decoding
audio_data, samplerate = sf.read(io.BytesIO(sample['audio']['bytes']))

Limitations

The content across modalities is matched based on general meaning, not exact identity. Consequently, the Image and Mesh components may differ in specific details. For instance, a caption of an image describing "Percy the Small Engine" might be paired with a 3D mesh of a standard green steam locomotive rather than the specific character from Thomas & Friends. This is because they share high semantic proximity in the EmbeddingGemma space.

Licensing Information

The Vividha dataset itself is released under the MIT License. However, users must respect the license of the source data:

• Objaverse: ODC-By 1.0

Citation

If you use this dataset in your research, please cite:

@misc{vinayumrethevividha2026,
  author = {Vinay Umrethe},
  title = {Vividha: The Four Forms of Intelligence},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/VINAY-UMRETHE/Vividha}}
}