Advances in Generative Models and Architectural Design

The field of generative models and architectural design is witnessing significant advancements, driven by innovations in autoregressive models, neural cellular automata, and transformer-based architectures. Researchers are exploring new paradigms, such as continuous latent space modeling and spatial-aware decay mechanisms, to improve the efficiency and quality of image and text generation. Notably, the development of novel frameworks like DisCon and Hita is enabling more effective capture of holistic relationships among token sequences and global image properties. These advancements have the potential to transform various applications, including intelligent architectural design, escape room puzzle generation, and high-resolution image synthesis.

Some noteworthy papers in this area include: RoomCraft, which proposes a multi-stage pipeline for generating coherent 3D indoor scenes from user inputs, demonstrating significant improvements in generating realistic and visually appealing room layouts. Neural Cellular Automata: From Cells to Pixels, which overcomes the limitation of low-resolution grids in neural cellular automata by pairing them with a tiny, shared implicit decoder, enabling the generation of full-HD outputs in real-time. Locality-aware Parallel Decoding for Efficient Autoregressive Image Generation, which accelerates autoregressive image generation through flexible parallelized autoregressive modeling and locality-aware generation ordering, achieving at least 3.4x lower latency than previous parallelized autoregressive models.

Sources

FloorPlan-DeepSeek (FPDS): A multimodal approach to floorplan generation using vector-based next room prediction

GenEscape: Hierarchical Multi-Agent Generation of Escape Room Puzzles

RoomCraft: Controllable and Complete 3D Indoor Scene Generation

Identification of Cellular Automata on Spaces of Bernoulli Probability Measures

Neural Cellular Automata: From Cells to Pixels

CycleVAR: Repurposing Autoregressive Model for Unsupervised One-Step Image Translation

Pipelined Decoder for Efficient Context-Aware Text Generation

Transition Matching: Scalable and Flexible Generative Modeling

Text-to-Level Diffusion Models With Various Text Encoders for Super Mario Bros

Flexible Language Modeling in Continuous Space with Transformer-based Autoregressive Flows

High-resolution spatial memory requires grid-cell-like neural codes

LEDOM: An Open and Fundamental Reverse Language Model

Autoregressive Image Generation with Linear Complexity: A Spatial-Aware Decay Perspective

Rethinking Discrete Tokens: Treating Them as Conditions for Continuous Autoregressive Image Synthesis

Locality-aware Parallel Decoding for Efficient Autoregressive Image Generation

Improving Constrained Generation in Language Models via Self-Distilled Twisted Sequential Monte Carlo

Holistic Tokenizer for Autoregressive Image Generation

Built with on top of