Conditional diffusion model for high quality unseen hologram synthesis.

Area of Science:

• Computer Vision
• Holography
• Artificial Intelligence

Background:

• Existing methods for unseen hologram synthesis struggle with physical inconsistencies and limited attribute control.
• Non-end-to-end models generate artifacts due to phase errors, while autoencoder-based methods lack attribute manipulation flexibility.

Purpose of the Study:

• To develop an end-to-end conditional diffusion model for direct unseen hologram synthesis from semantic and physical constraints.
• To enable precise control over holographic attributes like diffraction depth while minimizing model parameters.

Main Methods:

• Proposed holo-LDM, an end-to-end conditional diffusion model utilizing complex-valued UNets and cross-attention transformers.
• Employed physical model-driven UNets for generating speckle-suppressed holographic datasets.
• Integrated cross-attention transformers for dynamic label alignment with holographic features during denoising.

Main Results:

• Synthesized high-quality holograms across diverse categories with accurate diffraction depth control.
• Achieved over 30% improvement in Fréchet inception distance (FID) compared to non-end-to-end baselines.
• Reduced training costs by an order of magnitude with a compact 2 KB parameter model.

Conclusions:

• holo-LDM offers an efficient and effective solution for unseen hologram synthesis with fine-grained attribute control.
• The model demonstrates superior performance and reduced training overhead, paving the way for advanced mixed reality experiences.