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

Researcher-built Discrete Distribution Networks (DDN), accepted to ICLR 2025 with code released, introduce a new generative paradigm that models data with hierarchical discrete distributions. Instead of outputting one sample, each Discrete Distribution Layer (DDL) produces multiple discrete candidates; the sample closest to the ground truth (by L2) is selected and fed as a condition to the next layer. Layers stack to form a tree-structured latent space where depth exponentially expands representational capacity and samples map to leaf nodes. DDN can approximate continuous distributions by generating discrete sample points and refines outputs layer-by-layer. Practical training tricks like a Split-and-Prune strategy mitigate issues such as dead nodes and density shift, reducing KL divergence (reportedly below real-sample baselines for toy tasks). Experiments on CIFAR-10 and FFHQ (256×256) demonstrate reconstruction quality and the model’s unique 1D latent representation. DDN’s properties matter because they enable more general zero-shot conditional generation without gradient access—for example, text-to-image guidance via a black-box CLIP—plus reconstruction and apparent resistance to mode collapse. Memory overhead is slightly above a comparable GAN generator, but unchosen samples don’t keep gradients and inference can sample a single index to save compute. The author highlights fast development (≈3 months) and points to future work: scaling to ImageNet, hyperparameter/theory refinement, applying DDN to small-generation domains, non-pixel tasks, and language modeling.