# Meltdown: Circuits and Bifurcations in Point-Cloud-Conditioned 3D Diffusion Transformers arXiv: https://arxiv.org/abs/2602.11130 ยท HTML: https://mlsanity.com/paper/2602.11130/ ## Key facts - Cited by 1 papers in the map (data through 14 Oct 2026). - Placed in 3D, point cloud and scene generation, one of 14 directions in Diffusion models. - Builds on 28 of its 59 references that are in the map. - First posted on arXiv 11 Feb 2026. ## Placed in - Diffusion models / [3D, point cloud and scene generation](https://mlsanity.com/diffusion/3d-point-cloud-and-scene-generation/) ## Builds on (most cited first, 28 of 28) - [Denoising Diffusion Probabilistic Models](https://mlsanity.com/paper/2006.11239/) (2020) - [U-Net: Convolutional Networks for Biomedical Image Segmentation](https://mlsanity.com/paper/1505.04597/) (2015) - [High-Resolution Image Synthesis with Latent Diffusion Models](https://mlsanity.com/paper/2112.10752/) (2021) - [Classifier-Free Diffusion Guidance](https://mlsanity.com/paper/2207.12598/) (2022) - [PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation](https://mlsanity.com/paper/1612.00593/) (2016) - [Scaling Rectified Flow Transformers for High-Resolution Image Synthesis](https://mlsanity.com/paper/2403.03206/) (2024) - [Elucidating the Design Space of Diffusion-Based Generative Models](https://mlsanity.com/paper/2206.00364/) (2022) - [Towards Automated Circuit Discovery for Mechanistic Interpretability](https://mlsanity.com/paper/2304.14997/) (2023) - [Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items](https://mlsanity.com/paper/2204.11918/) (2022) - [Simple diffusion: End-to-end diffusion for high resolution images](https://mlsanity.com/paper/2301.11093/) (2023) - [Towards Best Practices of Activation Patching in Language Models: Metrics and Methods](https://mlsanity.com/paper/2309.16042/) (2023) - [What the DAAM: Interpreting Stable Diffusion Using Cross Attention](https://mlsanity.com/paper/2210.04885/) (2022) - [Inference-Time Scaling for Diffusion Models beyond Scaling Denoising Steps](https://mlsanity.com/paper/2501.09732/) (2025) - [How to use and interpret activation patching](https://mlsanity.com/paper/2404.15255/) (2024) - [Open Problems in Mechanistic Interpretability](https://mlsanity.com/paper/2501.16496/) (2025) - [PointCleanNet: Learning to Denoise and Remove Outliers from Dense Point Clouds](https://mlsanity.com/paper/1901.01060/) (2019) - [DiTAR: Diffusion Transformer Autoregressive Modeling for Speech Generation](https://mlsanity.com/paper/2502.03930/) (2025) - [Surface Reconstruction from Point Clouds: A Survey and a Benchmark](https://mlsanity.com/paper/2205.02413/) (2022) - [One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Models](https://mlsanity.com/paper/2410.22366/) (2024) - [Make-A-Shape: a Ten-Million-scale 3D Shape Model](https://mlsanity.com/paper/2401.11067/) (2024) - [Emergence and Evolution of Interpretable Concepts in Diffusion Models](https://mlsanity.com/paper/2504.15473/) (2025) - [Hunyuan3D-Omni: A Unified Framework for Controllable Generation of 3D Assets](https://mlsanity.com/paper/2509.21245/) (2025) - [A Geometric Perspective on Diffusion Models](https://mlsanity.com/paper/2305.19947/) (2023) - [The Information Dynamics of Generative Diffusion](https://mlsanity.com/paper/2508.19897/) (2025) - [A Survey and Benchmark of Automatic Surface Reconstruction from Point Clouds](https://mlsanity.com/paper/2301.13656/) (2023) - [Wavelet Latent Diffusion (Wala): Billion-Parameter 3D Generative Model with Compact Wavelet Encodings](https://mlsanity.com/paper/2411.08017/) (2024) - [Interpreting Large Text-to-Image Diffusion Models with Dictionary Learning](https://mlsanity.com/paper/2505.24360/) (2025) - [Deep Learning for 3D Point Cloud Enhancement: A Survey](https://mlsanity.com/paper/2411.00857/) (2024) ## Built on by (most cited first, 1 of 1) - [A Chosen Future Can Still Be Rewritten: Causal Writability in Video Models](https://mlsanity.com/paper/2609.15980/) (2026) ## Abstract (arXiv) Sparse point clouds are a common input modality for 3D surface reconstruction, including in safety-critical settings such as surgical navigation and autonomous perception. Recent point-cloud-conditioned 3D diffusion transformers achieve state-of-the-art results in this regime by leveraging learned priors. We show that these models can fail catastrophically under realistic input variation, and present a mechanistic case study of why. We identify a failure mode we call Meltdown: tiny on-surface perturbations to a sparse input point cloud can fracture the reconstructed output into hundreds of disconnected pieces. Adversarial search recovers Meltdown in 89.9-100% of shapes across the two open-weight state-of-the-art architectures we study (WaLa, Make-a-Shape) on real-world datasets (GSO, SimJEB) and under both DDPM and DDIM sampling. We trace Meltdown along the forward pass: it is governed by how uniformly the points are distributed on the surface, faithfully transduced through the point-cloud encoder, and committed by a single early-denoising cross-attention write in the diffusion backbone. Diffusion-trajectory ensembles exhibit symmetry-breaking near this commit step, consistent with a bifurcation of the reverse process. Through a suite of matched-magnitude controls, we show that the variable on which the model commits is directional, concentrated in a low-rank subspace of the write's perturbation drift. Motivated by this finding, we introduce PowerRemap, a test-time control that reshapes the singular spectrum of the localized write to suppress this drift, with rescue rates of 98.3% on WaLa and 84.6% on Make-a-Shape. Together, these results link a circuit-level cross-attention mechanism to a trajectory-level account of the failure, demonstrating how mechanistic analysis can explain and guide behavior in conditional diffusion transformers.