Flow3D-OPD: Multi-Teacher On-Policy Distillation for 3D Geometry Generation with Flow-Matching Diffusion Transformer

• URL: http://arxiv.org/abs/2609.07137v1
• Date: Mon, 07 Sep 2026 07:36:52 GMT
• Title: Flow3D-OPD: Multi-Teacher On-Policy Distillation for 3D Geometry Generation with Flow-Matching Diffusion Transformer
• Abstract summary: We propose textbfFlow3D-OPD, a two-stage post-training framework for 3D geometry generation.<n>In the first stage, we utilize the semi-policy to enhance the foundational capability of the pretrained model.<n>In the second stage, we consolidate heterogeneous expertise into a unified student model.
• Score: 26.474560149220917
• License: http://creativecommons.org/licenses/by-nc-sa/4.0/
• Abstract: Recent image-to-3D generation models built on flow-matching diffusion Transformers (DiT) can produce high-fidelity meshes, yet their post-training strategy remains largely unexplored. There exist several critical bottlenecks in reinforcement learning: the inherent difficulty of defining comprehensive rewards for 3D geometric quality, and the gradient interference that arises when jointly optimizing heterogeneous objectives. Inspired by the practicability of on-policy distillation (OPD) in large language models and image generation, we propose \textbf{Flow3D-OPD}, a two-stage post-training framework that introduces multi-teacher distillation into 3D geometry generation. In the first stage, we utilize the semi-policy to enhance the foundational capability of the pretrained model and then design an agentic verifier for 3D geometric quality evaluation. Based on the verifier, we could cultivate domain-specialized teacher models via direct preference optimization (DPO). In the second stage, we consolidate heterogeneous expertise into a unified student model through on-policy distillation with hard task-routing sampling and gradient accumulation, which could mitigate the gradient interference in joint optimization. Without relying on elaborate modifications, our straightforward yet effective design achieves consistent improvements across all geometric quality dimensions and surpasses all teacher models in the average metric. Extensive experiments demonstrate that our approach provides an effective paradigm for reinforcement learning in 3D generation.

Related papers

• FlowObject: Flow Steering for Bridging Generative Priors and Reconstruction Fidelity [34.142288038459704]
  FlowObject is a framework that reformulates sparse-view 3D reconstruction as a training-free, guided inverse problem.<n>Our method significantly outperforms state-of-the-art generative models and optimization-based frameworks in both geometric completeness and view-dependent appearance fidelity.
  arXiv  Detail & Related papers  (2026-06-17T12:42:09Z)
• GTA: Advancing Image-to-3D World Generation via Geometry Then Appearance Video Diffusion [29.999238067855245]
  We propose GTA, a novel image-to-3D world generation method following a Geometry-Then-Appearance paradigm.<n>Specifically, given a single input image, GTA adopts a two-stage framework with two dedicated video diffusion models.<n>Extensive experiments have demonstrated that our proposed method consistently outperforms existing approaches in terms of fidelity, visual quality, and geometric accuracy.
  arXiv  Detail & Related papers  (2026-05-13T03:43:02Z)
• UniPre3D: Unified Pre-training of 3D Point Cloud Models with Cross-Modal Gaussian Splatting [64.31900521467362]
  No existing pre-training method is equally effective for both object- and scene-level point clouds.<n>We introduce UniPre3D, the first unified pre-training method that can be seamlessly applied to point clouds of any scale and 3D models of any architecture.
  arXiv  Detail & Related papers  (2025-06-11T17:23:21Z)
• Bridging Geometry-Coherent Text-to-3D Generation with Multi-View Diffusion Priors and Gaussian Splatting [51.08718483081347]
  We propose a framework that couples multi-view joint distribution priors to ensure geometrically consistent 3D generation.<n>We derive an effective optimization rule that effectively couples multi-view priors to guide optimization across different viewpoints.<n>We employ a deformable tetrahedral grid, from 3D-GS and refined through CSD, to produce high-quality, refined meshes.
  arXiv  Detail & Related papers  (2025-05-07T09:12:45Z)
• TripoSG: High-Fidelity 3D Shape Synthesis using Large-Scale Rectified Flow Models [69.0220314849478]
  TripoSG is a new streamlined shape diffusion paradigm capable of generating high-fidelity 3D meshes with precise correspondence to input images.<n>The resulting 3D shapes exhibit enhanced detail due to high-resolution capabilities and demonstrate exceptional fidelity to input images.<n>To foster progress and innovation in the field of 3D generation, we will make our model publicly available.
  arXiv  Detail & Related papers  (2025-02-10T16:07:54Z)
• DreamPolish: Domain Score Distillation With Progressive Geometry Generation [66.94803919328815]
  We introduce DreamPolish, a text-to-3D generation model that excels in producing refined geometry and high-quality textures. In the geometry construction phase, our approach leverages multiple neural representations to enhance the stability of the synthesis process. In the texture generation phase, we introduce a novel score distillation objective, namely domain score distillation (DSD), to guide neural representations toward such a domain.
  arXiv  Detail & Related papers  (2024-11-03T15:15:01Z)
• Masked Generative Extractor for Synergistic Representation and 3D Generation of Point Clouds [6.69660410213287]
  We propose an innovative framework called Point-MGE to explore the benefits of deeply integrating 3D representation learning and generative learning. In shape classification, Point-MGE achieved an accuracy of 94.2% (+1.0%) on the ModelNet40 dataset and 92.9% (+5.5%) on the ScanObjectNN dataset. Experimental results also confirmed that Point-MGE can generate high-quality 3D shapes in both unconditional and conditional settings.
  arXiv  Detail & Related papers  (2024-06-25T07:57:03Z)
• Pushing Auto-regressive Models for 3D Shape Generation at Capacity and Scalability [118.26563926533517]
  Auto-regressive models have achieved impressive results in 2D image generation by modeling joint distributions in grid space. We extend auto-regressive models to 3D domains, and seek a stronger ability of 3D shape generation by improving auto-regressive models at capacity and scalability simultaneously.
  arXiv  Detail & Related papers  (2024-02-19T15:33:09Z)
• Learning Versatile 3D Shape Generation with Improved AR Models [91.87115744375052]
  Auto-regressive (AR) models have achieved impressive results in 2D image generation by modeling joint distributions in the grid space. We propose the Improved Auto-regressive Model (ImAM) for 3D shape generation, which applies discrete representation learning based on a latent vector instead of volumetric grids.
  arXiv  Detail & Related papers  (2023-03-26T12:03:18Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.