INRCT: An End-to-End Framework for Encoding and Generating Implicit Neural Representation

ICLR 2026 Conference Submission16468 Authors

Keywords: Implicit Neural Representation, Diffusion Models, Consistency Training

TL;DR: We propose INRCT, a unified and end-to-end training generative framework for modality-agnostic INR modeling, supporting both INR encoding and generation in a single model.

Abstract: Current diffusion models based on implicit neural representations (INRs) typically adopt a two-stage framework: an encoder is first trained to map signals into a latent INR space, followed by a diffusion model that generates latent codes from noise. This design requires training and maintaining two separate models, introducing compounded reconstruction errors through the latent-to-data mapping and often leading to increased system complexity. In this work, we propose INRCT, a unified and end-to-end training generative framework for modality-agnostic INR modeling. Instead of operating in the latent space, INRCT performs diffusion directly in the data space by training a single INR hyper-network as a denoiser. Given noisy observations at different noise levels, INRCT predicts the INR for the corresponding clean signal, which is then rendered into data space for supervision. Our training objective coherently integrates a generation loss and a reconstruction loss to jointly support INR generation from noise and INR encoding from real signals within a single model. Extensive experiments on multiple benchmark datasets demonstrate that INRCT achieves superior generation and reconstruction performance compared to existing two-stage generative INR methods, while significantly improving the model efficiency and simplifying model design.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 16468