XH

X. Han

info

Please Note

2 records found

Master thesis (2026) - X. Han, P. Kellnhofer, X. Zhang
Current text-to-3D generation frameworks are bottlenecked by extreme computational demands, relying heavily on expensive test-time optimization or massive datasets to learn multi-modal alignments. This research investigates exploiting the highly structured semantic latent space of pre-trained Large Language Models (LLMs) to directly predict continuous 3D shapes, bypassing the need to build intricate alignment networks from scratch. To bridge the topological mismatch between language and 3D space, we propose a decoupled, two-stage architecture utilizing continuous Signed Distance Functions (SDFs). First, an auto-decoder optimization framework explicitly aligns the semantic text space with the geometric shape space to ensure high structural fidelity. Second, a mapping network learns the non-linear projection between the frozen LLM text latents and the optimized geometric latents. Extensive evaluations demonstrate that this decoupled paradigm prevents the geometric collapse observed in fully joint, end-to-end training, and confirms that pre-trained LLM embeddings serve as superior geometric priors compared to random initializations. Ultimately, this methodology establishes a stable, optimization-free inference pipeline, enabling feedforward synthesis of 3D geometries directly from raw textual prompts. ...
Estimating reverberation time (RT60) accurately is crucial for enhancing the acoustic quality of various environments as it decides how you feel the sound fades away subjectively. Traditional methods, such as Sabine's equation, require extensive prior knowledge and assume ideal conditions, limiting their practicality. To address these limitations, this paper explores the application of convolutional neural networks (CNNs) enhanced with an encoder architecture based on transformer mechanisms for blind RT60 estimation. The proposed model leverages simulated and real-world datasets, incorporating environmental noise to improve robustness. Results indicate that the CNN-Encoder model achieves superior performance, with a mean squared error (MSE) as low as 0.0006 seconds for pure room impulse responses (RIRs) and 0.0011 seconds under +30dB signal-to-noise ratio (SNR) conditions. It also demonstrates potential in practical usage achieving an MSE of 0.0282 seconds under audio recordings. This approach offers a significant reduction in estimation error compared to the CNN-only architecture, demonstrating the potential for improved acoustic parameter estimation in varied environments. Future work will focus on further optimizing the model for real-world applications and reducing computational complexity while maintaining high accuracy. ...