Krishna Sri Ipsit Mantri

Bo Peng, Eric Alcaide, Quentin Anthony et al.

Transformers have revolutionized almost all natural language processing (NLP) tasks but suffer from memory and computational complexity that scales quadratically with sequence length. In contrast, recurrent neural networks (RNNs) exhibit linear scaling in memory and computational requirements but struggle to match the same performance as Transformers due to limitations in parallelization and scalability. We propose a novel model architecture, Receptance Weighted Key Value (RWKV), that combines the efficient parallelizable training of transformers with the efficient inference of RNNs. Our approach leverages a linear attention mechanism and allows us to formulate the model as either a Transformer or an RNN, thus parallelizing computations during training and maintains constant computational and memory complexity during inference. We scale our models as large as 14 billion parameters, by far the largest dense RNN ever trained, and find RWKV performs on par with similarly sized Transformers, suggesting future work can leverage this architecture to create more efficient models. This work presents a significant step towards reconciling trade-offs between computational efficiency and model performance in sequence processing tasks.

Krishna Sri Ipsit Mantri, Nevasini Sasikumar

Fashionable image generation aims to synthesize images of diverse fashion prevalent around the globe, helping fashion designers in real-time visualization by giving them a basic customized structure of how a specific design preference would look in real life and what further improvements can be made for enhanced customer satisfaction. Moreover, users can alone interact and generate fashionable images by just giving a few simple prompts. Recently, diffusion models have gained popularity as generative models owing to their flexibility and generation of realistic images from Gaussian noise. Latent diffusion models are a type of generative model that use diffusion processes to model the generation of complex data, such as images, audio, or text. They are called "latent" because they learn a hidden representation, or latent variable, of the data that captures its underlying structure. We propose a method exploiting the equivalence between diffusion models and energy-based models (EBMs) and suggesting ways to compose multiple probability distributions. We describe a pipeline on how our method can be used specifically for new fashionable outfit generation and virtual try-on using LLM-guided text-to-image generation. Our results indicate that using an LLM to refine the prompts to the latent diffusion model assists in generating globally creative and culturally diversified fashion styles and reducing bias.