arXiv Endorsement Request: From Weight-Space Diffusion to Latent-Space DeepSDF - [cs.CV / cs.GR]

Hello community, I am seeking an arXiv endorsement for a recent research project investigating image-to-3D generation. Due to the updated 2026 submission policies, I am looking for a peer review and endorsement from the community. Paper Title: From Weight-Space Diffusion to Latent-Space DeepSDF: An Empirical Investigation of Image-Conditioned 3-D Generation at Small Data Scale Target Categories: * Primary: Computer Vision and Pattern Recognition (cs.CV) * Secondary: Graphics (cs.GR); Machine Learning (cs.LG) Abstract: We present a systematic investigation of 3D shape generation under severe data constraints (≤976 shapes). The work traces the structural failure modes of 54,785-dimensional weight-space diffusion—specifically the “warm-start dominance” problem (0.96 mean cosine similarity)—and demonstrates a successful pivot to a DINOv2-conditioned Latent Diffusion Model (LDM) using DeepSDF embeddings. Our results show that architectural inductive biases are more critical for out-of-distribution generalization than learned compression at this data scale. Link to PDF: https://jainaditya.in/whitepaper/hypernet-deepsdf Researcher Details: * Author: Aaditya Jain * ORCID: 0009-0005-5534-5641 * Affiliation: Independent Thesis Research I have documented the twelve experimental phases, including the failure of weight-space autoencoders and the success of the latent DiT pipeline. I am happy to provide my arXiv ID or discuss the technical logs (Phase 1-12) with anyone willing to review the work for endorsement. Thank you for supporting independent research in 3D machine learning!