Diffusion models have achieved demonstrated huge success in generative tasks across images, audio, and text. However, applying them to modality translation — converting data from one modality to another (e.g., images to 3D shapes, low-res to high-res) — remains limited by assumptions like matching dimensionality or architecture-specific designs.

We introduce LDDBM, a general-purpose framework for modality translation using a latent extension of Denoising Diffusion Bridge Models (DDBMs). Our method operates in a shared latent space, avoids restrictive assumptions, and introduces two key innovations:

• A contrastive loss that enforces semantic alignment across modalities

• A predictive loss that directly improves translation quality

Our model performs well across diverse tasks — including multi-view to 3D generation, zero-shot image super-resolution, and scene occupancy prediction — establishing a new state-of-the-art for general modality translation.

We extend DDBMs into a latent diffusion bridge, enabling translation between modalities of different shapes and semantics. The architecture consists of:

• Modality-specific encoders and decoders

• A shared latent space where translation occurs via a learned diffusion bridge

• An encoder-decoder Transformer designed specifically for bridging heterogeneous representations

Our training loss combines:

1. A bridge loss (score matching)

2. A predictive loss to reconstruct the target

3. A contrastive loss (InfoNCE) to align latent pairs