Show HN: I invented a new generative model and got accepted to ICLR (discrete-distribution-networks.github.io)

Article summary

The article introduces a novel generative model called Discrete Distribution Networks (DDN), which approximates data distribution using hierarchical discrete distributions. DDN has been accepted to ICLR 2025 and has shown promising results in image generation and zero-shot conditional generation tasks. The model uses a unique architecture that generates multiple samples simultaneously and selects the one closest to the target distribution. The author discusses potential future research directions, including scaling up the model and applying it to non-image domains.

Main themes

• Generative Models
• Discrete Distribution Networks
• Image Generation
• Zero-Shot Conditional Generation
• Model Architecture
• Future Research Directions

What commenters say

• The DDN model has the potential to be applied to non-image domains, such as language modeling and object detection, due to its unique architecture and properties.
• The model's ability to generate multiple samples simultaneously and select the closest one to the target distribution makes it efficient and effective for certain tasks.
• Some commentators question the lack of baseline comparisons in the paper, which makes it difficult to evaluate the model's performance relative to existing methods.
• The DDN model is seen as a promising alternative to diffusion models, with advantages such as single forward pass generation and ease of imposing constraints during generation.
• The model's hierarchical generation and tree-structured latent space are novel and exciting features that could lead to new applications and research directions.
• There are concerns about the model's scalability and ability to handle high-dimensional target distributions, which could limit its applicability in certain domains.
• The author's decision to publish the paper despite limited experimentation and analysis is defended as a way to share novel ideas and stimulate further research, even if the results are not yet state-of-the-art.