pith. sign in

arxiv: 2601.06928 · v2 · submitted 2026-01-11 · 💻 cs.CV

RenderFlow: Single-Step Neural Rendering via Flow Matching

Shenghao Zhang , Runtao Liu , Christopher Schroers , Yang Zhang This is my paper

Pith reviewed 2026-05-16 14:57 UTC · model grok-4.3

classification 💻 cs.CV
keywords neural renderingflow matchingsingle-step renderingG-buffersphotorealistic renderinginverse renderingtemporal consistencyreal-time graphics
0
0 comments X

The pith

Flow matching enables single-step deterministic neural rendering from geometry buffers that reaches near real-time photorealistic quality.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to replace the slow iterative sampling and random noise of diffusion-based neural renderers with a single forward pass through a flow-matching network. The model ingests geometry buffers and can take optional sparse keyframes as guidance to improve consistency and physical accuracy. If the approach works, rendering pipelines could run fast enough for interactive use while still producing images that hold up against traditional light-transport simulation. The same pretrained network can be lightly adapted to perform inverse tasks such as breaking an image into albedo, normals, and other intrinsic layers.

Core claim

RenderFlow is an end-to-end deterministic single-step neural rendering framework built on a flow matching paradigm. An efficient sparse keyframe guidance module further strengthens rendering quality and generalization. The resulting pipeline reaches near real-time speeds with photorealistic output and can be repurposed, via a lightweight adapter, for inverse rendering tasks such as intrinsic decomposition.

What carries the argument

Flow matching paradigm that maps geometry buffers to images in one deterministic step, augmented by a sparse keyframe guidance module.

Load-bearing premise

That a flow-matching model trained on geometry buffers and optional sparse keyframes can produce outputs that are simultaneously photorealistic, physically plausible, and temporally consistent without the stochasticity of diffusion processes.

What would settle it

Run the model on a held-out scene with known ground-truth physically based renders and measure whether perceptual error stays below a small threshold while a moving-camera sequence shows no visible temporal flicker or inconsistency.

Figures

Figures reproduced from arXiv: 2601.06928 by Christopher Schroers, Runtao Liu, Shenghao Zhang, Yang Zhang.

Figure 1
Figure 1. Figure 1: Overview of the proposed RenderFlow framework. Left: RenderFlow learns a single-step conditional flow from albedo (rather than noise) to shaded images, yielding ∼10× faster rendering while better physical light transport; both forward and inverse rendering are unified within a single model. Right: Example visualizations and corresponding error maps illustrating fidelity. Abstract Conventional physically ba… view at source ↗
Figure 2
Figure 2. Figure 2: An overview of our proposed RenderFlow architecture. Unlike diffusion methods that start from noise, RenderFlow takes albedo as input and directly predicts fully-shaded outputs. Built on a pre-trained video DiT, it injects aligned G-buffer tokens into trans￾former blocks. Environment maps and optional keyframes are integrated via lightweight adapters (see [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Design of the customized transformer block in RenderFlow. To incorporate global lighting context, the environment map is injected via an adaptive normalization layer before the transformer layers To compensate for the limited information in per-frame G￾buffers, the keyframe adapter introduces cross-attention to extract complementary temporal features from an optional keyframe. The adapter includes a LoRA m… view at source ↗
Figure 4
Figure 4. Figure 4: Inverse adapter architecture. The inverse adapter uses the frozen forward rendering backbone and introduces specialized modules to adapt it for G-buffer decomposition. into a latent zrgb, which is then mapped to tokens by a train￾able inverse embedder. To repurpose the frozen transformer for intrinsic decomposition, we augment its self-attention projections with low-rank adaptation (LoRA) modules and inser… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of rendered images with traditional methods including deferred rendering, and diffusion-based approaches (RGB-X and DiffusionRenderer). Our method better preserves texture details and produces high-quality shadows and lighting effects. collection of assets, significantly increasing data diver￾sity. This collection includes approximately 4,000 unique meshes and 30 high-dynamic-range i… view at source ↗
Figure 6
Figure 6. Figure 6: Metrics are evaluated over a 100-frame sequence and averaged across 10 runs. Shaded regions indicate variance across runs. Our method not only achieves the best overall perfor￾mance but also zero variance, reflecting its deterministic, consis￾tently high-quality behavior in contrast to the stochastic baselines. dard image reconstruction metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Inde… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of intrinsic decomposition. We compare our adapter-based method against RGB-X and DiffusionRen￾derer. This figure demonstrates the decomposition results for key G-buffer components: Albedo, Normal, Metallic, and Roughness. Method Paradigm Params PSNR ↑ SSIM ↑ LPIPS ↓ Time (s) ↓ Path Tracing Traditional - - - - >10 IBL Traditional - 16.903 0.882 0.105 - Deferred Traditional - 24.649 0… view at source ↗
Figure 8
Figure 8. Figure 8: Two different keyframe adapter designs. The left one reuses the query from the self-attention layer for efficiency, while the right one uses a dedicated query. or in a uniform timesteps. For the SDE-based approach, we follow LBM [16] and set the bridge noise σ to 0.005. As shown in Tab. 3, we empirically find that single-step inference yields superior results for models trained with multi-step schedules, w… view at source ↗
Figure 9
Figure 9. Figure 9: Demonstration of controlled material editing. Our model is capable of synthesizing smooth transitions for individual material properties while all other scene parameters are held constant. From left to right: the cone’s roughness is interpolated from 1.0 to 0.0; the cylinder’s metallic property from 0.0 to 1.0; and the cube’s diffuse albedo from blue to green. fixed keyframe gap of 16 frames, the adapter’s… view at source ↗
Figure 11
Figure 11. Figure 11: Analysis on inference steps. We show that multi-step inference can generate sharper shadows. However, due to mis￾alignments with ground truth, it will lead to lower metrics (PSNR). . We tested the inference steps without using keyframes. model is trained to predict the final rendered latent zˆ1 di￾rectly from the initial albedo latent z0. Due to the inaccu￾racy of each network evaluation, intermediate err… view at source ↗
Figure 10
Figure 10. Figure 10: A preview of our synthesized dataset, showcasing a [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Failed cases and analysis. Top: Rendered images for scenes with highly complex geometries. Bottom: The VAE compression leads to loss of fine-grained details Frame 0 Frame 1 Frame 2 Frame 3 Frame 4 [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Analysis on video latent downsampling. The causal design of the VAE encoder leads to the phenomenon that the initial frame retains sharpness while subsequent frames exhibit progres￾sive blurring due to temporal downsampling dependencies. generate these high-frequency details with greater physical plausibility and accuracy. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Additional qualitative forward rendering results. We provide more qualitative comparisons with error maps to highlight the differences between our predictions and the ground truth. For keyframe guidance, we sample keyframes with a 25-frame gap. Our method clearly outperforms the baseline methods. In addition, with the keyframe guidance our model achieves the best results among all models. 15 [PITH_FULL_I… view at source ↗
Figure 15
Figure 15. Figure 15: Additional qualitative forward rendering results (continued). 16 [PITH_FULL_IMAGE:figures/full_fig_p016_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Additional qualitative inverse rendering results. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_16.png] view at source ↗
read the original abstract

Conventional physically based rendering (PBR) pipelines generate photorealistic images through computationally intensive light transport simulations. Although recent deep learning approaches leverage diffusion model priors with geometry buffers (G-buffers) to produce visually compelling results without explicit scene geometry or light simulation, they remain constrained by two major limitations. First, the iterative nature of the diffusion process introduces substantial latency. Second, the inherent stochasticity of these generative models compromises physical accuracy and temporal consistency. In response to these challenges, we propose a novel, end-to-end, deterministic, single-step neural rendering framework, RenderFlow, built upon a flow matching paradigm. To further strengthen both rendering quality and generalization, we propose an efficient and effective module for sparse keyframe guidance. Our method significantly accelerates the rendering process and, by optionally incorporating sparsely rendered keyframes as guidance, enhances both the physical plausibility and overall visual quality of the output. The resulting pipeline achieves near real-time performance with photorealistic rendering quality, effectively bridging the gap between the efficiency of modern generative models and the precision of traditional physically based rendering. Furthermore, we demonstrate the versatility of our framework by introducing a lightweight, adapter-based module that efficiently repurposes the pretrained forward model for the inverse rendering task of intrinsic decomposition.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes RenderFlow, a deterministic single-step neural rendering framework based on flow matching. Conditioned on geometry buffers (G-buffers) and optional sparse rendered keyframes, the method aims to produce photorealistic images at near real-time speeds while improving physical plausibility and temporal consistency over stochastic diffusion approaches. It further introduces a lightweight adapter module to repurpose the pretrained model for inverse rendering tasks such as intrinsic decomposition.

Significance. If the empirical claims hold, the work would be significant for real-time graphics and vision applications by offering a faster, deterministic bridge between generative neural renderers and traditional physically based rendering. The sparse keyframe guidance and adapter for inverse tasks are practical additions that could improve generalization and enable new workflows without retraining from scratch.

major comments (2)
  1. [Abstract] Abstract: the claim that sparse keyframe guidance 'enhances both the physical plausibility and overall visual quality' is load-bearing for the central contribution, yet the abstract provides no description of how the flow-matching velocity field or loss incorporates explicit light-transport constraints (energy conservation, view-consistent BRDFs, or radiative transfer); without such mechanisms the outputs remain purely statistical and may fail on novel lighting or materials.
  2. [Abstract] Abstract: no quantitative results, error metrics, baselines, ablation studies, or runtime measurements are referenced to support the assertions of 'near real-time performance' and 'photorealistic rendering quality'; this absence prevents assessment of whether the single-step deterministic pipeline actually closes the gap with PBR precision.
minor comments (1)
  1. The abstract would benefit from a brief statement of the training dataset, resolution, and hardware used for the reported near-real-time performance.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract. We address each comment below and will revise the manuscript accordingly to strengthen clarity and support for our claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that sparse keyframe guidance 'enhances both the physical plausibility and overall visual quality' is load-bearing for the central contribution, yet the abstract provides no description of how the flow-matching velocity field or loss incorporates explicit light-transport constraints (energy conservation, view-consistent BRDFs, or radiative transfer); without such mechanisms the outputs remain purely statistical and may fail on novel lighting or materials.

    Authors: We appreciate this observation. RenderFlow is trained end-to-end on image pairs generated by a physically based renderer, so the learned velocity field implicitly encodes light-transport statistics present in the training data. The sparse keyframe guidance module conditions the flow on additional G-buffer and radiance information rendered from the same PBR pipeline at selected viewpoints; this conditioning injects explicit physical observations into the generation process without requiring hand-crafted constraints inside the velocity field or loss. We acknowledge that this remains a data-driven rather than explicitly constrained approach and will revise the abstract to clarify the mechanism and note the reliance on PBR-generated training data. revision: yes

  2. Referee: [Abstract] Abstract: no quantitative results, error metrics, baselines, ablation studies, or runtime measurements are referenced to support the assertions of 'near real-time performance' and 'photorealistic rendering quality'; this absence prevents assessment of whether the single-step deterministic pipeline actually closes the gap with PBR precision.

    Authors: We agree that the abstract would benefit from explicit quantitative anchors. In the revised version we will add concise references to the key metrics reported in Section 4 (e.g., PSNR/SSIM gains over diffusion baselines, runtime in milliseconds on consumer GPUs, and ablation results on keyframe density) while remaining within the abstract length limit. revision: yes

Circularity Check

0 steps flagged

No significant circularity in RenderFlow derivation

full rationale

The paper introduces RenderFlow as a novel deterministic single-step neural rendering framework built on the established external flow-matching paradigm, augmented by an optional sparse-keyframe guidance module. The provided abstract and description contain no equations, no fitted parameters relabeled as predictions, no self-citations invoked for uniqueness or load-bearing premises, and no ansatzes or known results renamed as new derivations. All central claims (near-real-time performance, improved physical plausibility via guidance) are presented as consequences of the architectural choices rather than reductions to the inputs by construction. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; full paper would be required to audit training objectives or architectural assumptions.

pith-pipeline@v0.9.0 · 5519 in / 1024 out tokens · 47997 ms · 2026-05-16T14:57:39.616886+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. D-Rex : Diffusion Rendering for Relightable Expressive Avatars

    cs.GR 2026-04 conditional novelty 7.0

    D-Rex applies a LoRA-fine-tuned video diffusion model as an image-space post-process to add consistent relighting to any expressive full-body avatar pipeline while preserving motion and facial detail.

Reference graph

Works this paper leans on

42 extracted references · 42 canonical work pages · cited by 1 Pith paper · 6 internal anchors

  1. [1]

    Attila T. ´Afra. Intel ® Open Image Denoise, 2025.https: //www.openimagedenoise.org. 6, 14

    work page 2025

  2. [2]

    Stochastic Interpolants: A Unifying Framework for Flows and Diffusions

    Michael S Albergo, Nicholas M Boffi, and Eric Vanden- Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions.arXiv preprint arXiv:2303.08797,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Physically-based shading at disney

    Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. InAcm siggraph, pages 1–7. vol. 2012, 2012. 3

    work page 2012

  4. [4]

    Flash diffusion: Accelerating any conditional diffusion model for few steps image generation

    Clement Chadebec, Onur Tasar, Eyal Benaroche, and Ben- jamin Aubin. Flash diffusion: Accelerating any conditional diffusion model for few steps image generation. InProceed- ings of the AAAI Conference on Artificial Intelligence, pages 15686–15695, 2025. 7

    work page 2025

  5. [5]

    Lbm: Latent bridge match- ing for fast image-to-image translation.arXiv preprint arXiv:2503.07535, 2025

    Cl ´ement Chadebec, Onur Tasar, Sanjeev Sreetharan, and Benjamin Aubin. Lbm: Latent bridge match- ing for fast image-to-image translation.arXiv preprint arXiv:2503.07535, 2025. 2, 5

    work page arXiv 2025

  6. [6]

    Scaling recti- fied flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Muller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. InForty-first international conference on machine learning,

    work page

  7. [7]

    arXiv preprint arXiv:2506.15673 , year=

    Kai He, Ruofan Liang, Jacob Munkberg, Jon Hasselgren, Nandita Vijaykumar, Alexander Keller, Sanja Fidler, Igor Gilitschenski, Zan Gojcic, and Zian Wang. Unirelight: Learning joint decomposition and synthesis for video relight- ing.arXiv preprint arXiv:2506.15673, 2025. 2

    work page arXiv 2025

  8. [8]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 3

    work page 2020

  9. [9]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022. 8

    work page 2022

  10. [10]

    Ak Peters Natick, 2001

    Henrik Wann Jensen.Realistic image synthesis using photon mapping. Ak Peters Natick, 2001. 1

    work page 2001

  11. [11]

    Res-tuning: A flexible and efficient tuning paradigm via unbinding tuner from backbone.Advances in Neural Information Processing Systems, 36:42689–42716, 2023

    Zeyinzi Jiang, Chaojie Mao, Ziyuan Huang, Ao Ma, Yiliang Lv, Yujun Shen, Deli Zhao, and Jingren Zhou. Res-tuning: A flexible and efficient tuning paradigm via unbinding tuner from backbone.Advances in Neural Information Processing Systems, 36:42689–42716, 2023. 8

    work page 2023

  12. [12]

    VACE: All-in-One Video Creation and Editing

    Zeyinzi Jiang, Zhen Han, Chaojie Mao, Jingfeng Zhang, Yulin Pan, and Yu Liu. Vace: All-in-one video creation and editing.arXiv preprint arXiv:2503.07598, 2025. 4, 5, 11

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  13. [13]

    Real Shading in Unreal Engine 4

    Brian Karis. Real Shading in Unreal Engine 4. 2013. 3

    work page 2013

  14. [14]

    Flux.1 kontext: Flow matching for in-context image generation and editing in latent space,

    Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dock- horn, Jack English, Zion English, Patrick Esser, Sumith Ku- lal, Kyle Lacey, Yam Levi, Cheng Li, Dominik Lorenz, Jonas M¨uller, Dustin Podell, Robin Rombach, Harry Saini, Axel Sauer, and Luke Smith. Flux.1 kontext: Flow matching for in-context i...

    work page

  15. [15]

    Diffusion- renderer: Neural inverse and forward rendering with video diffusion models

    Ruofan Liang, Zan Gojcic, Huan Ling, Jacob Munkberg, Jon Hasselgren, Zhi-Hao Lin, Jun Gao, Alexander Keller, Nan- dita Vijaykumar, Sanja Fidler, and Zian Wang. Diffusion- renderer: Neural inverse and forward rendering with video diffusion models. InThe IEEE Conference on Computer Vi- sion and Pattern Recognition (CVPR), 2025. 2, 4, 6, 7, 14

    work page 2025

  16. [16]

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022. 2, 3, 8

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  17. [17]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017. 11

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  18. [18]

    Structure-preserving super resolution with gradient guidance

    Cheng Ma, Yongming Rao, Yean Cheng, Ce Chen, Jiwen Lu, and Jie Zhou. Structure-preserving super resolution with gradient guidance. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 7769–7778, 2020. 5

    work page 2020

  19. [19]

    Deep shading: convolutional neural networks for screen space shading

    Oliver Nalbach, Elena Arabadzhiyska, Dushyant Mehta, H- P Seidel, and Tobias Ritschel. Deep shading: convolutional neural networks for screen space shading. InComputer graphics forum, pages 65–78. Wiley Online Library, 2017. 2

    work page 2017

  20. [20]

    One-Step Image Translation with Text-to- Image Models, 2024

    Gaurav Parmar, Taesung Park, Srinivasa Narasimhan, and Jun-Yan Zhu. One-Step Image Translation with Text-to- Image Models, 2024. 2, 5

    work page 2024

  21. [21]

    MIT Press, 2023

    Matt Pharr, Wenzel Jakob, and Greg Humphreys.Physi- cally based rendering: From theory to implementation. MIT Press, 2023. 1

    work page 2023

  22. [22]

    Poly Haven: The public 3d asset library

    Poly Haven. Poly Haven: The public 3d asset library. https://polyhaven.com/, 2025. Accessed: August 3, 2025. 14

    work page 2025

  23. [23]

    Ren ´e Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer.IEEE transactions on pattern analysis and machine intelligence, 44(3):1623–1637, 2020. 5, 11

    work page 2020

  24. [24]

    Photographic tone reproduction for digital images

    Erik Reinhard, Michael Stark, Peter Shirley, and James Fer- werda. Photographic tone reproduction for digital images. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 661–670. 2023. 4

    work page 2023

  25. [25]

    Lightformer: Light-oriented global neural rendering in dynamic scene.ACM Transactions on Graph- ics, 43(4):1–14, 2024

    Haocheng Ren, Yuchi Huo, Yifan Peng, Hongtao Sheng, Hongxiang Huang, Weidong Xue, Jingzhen Lan, Rui Wang, and Hujun Bao. Lightformer: Light-oriented global neural rendering in dynamic scene.ACM Transactions on Graph- ics, 43(4):1–14, 2024. 2

    work page 2024

  26. [26]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1

    work page 2022

  27. [27]

    Fast high- resolution image synthesis with latent adversarial diffusion distillation

    Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high- resolution image synthesis with latent adversarial diffusion distillation. InSIGGRAPH Asia 2024 Conference Papers, pages 1–11, 2024. 7

    work page 2024

  28. [28]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. InEuropean 9 Conference on Computer Vision, pages 87–103. Springer,

    work page

  29. [29]

    Diffusion schr ¨odinger bridge matching

    Yuyang Shi, Valentin De Bortoli, Andrew Campbell, and Arnaud Doucet. Diffusion schr ¨odinger bridge matching. Advances in Neural Information Processing Systems, 36: 62183–62223, 2023. 3

    work page 2023

  30. [30]

    Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063,

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neurocomputing, 568:127063,

    work page

  31. [31]

    Deep Illumination: Approximating Dynamic Global Illumination with Generative Adversarial Network

    Manu Mathew Thomas and Angus G Forbes. Deep illumination: Approximating dynamic global illumina- tion with generative adversarial network.arXiv preprint arXiv:1710.09834, 2017. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  32. [32]

    Unreal engine marketplace.https:// www.fab.com/channels/unreal- engine, 2025

    Unreal Engine. Unreal engine marketplace.https:// www.fab.com/channels/unreal- engine, 2025. Accessed: August 3, 2025. 13

    work page 2025

  33. [33]

    Wan: Open and Advanced Large-Scale Video Generative Models

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianx- iao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jin- gren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pandeng Li, Pingyu Wu, Ruihang Chu, Ruili Feng, Shiwei Zhang, Siyang Sun, Tao Fan...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  34. [34]

    Difix3d+: Improving 3d reconstruc- tions with single-step diffusion models

    Jay Zhangjie Wu, Yuxuan Zhang, Haithem Turki, Xuanchi Ren, Jun Gao, Mike Zheng Shou, Sanja Fidler, Zan Goj- cic, and Huan Ling. Difix3d+: Improving 3d reconstruc- tions with single-step diffusion models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 26024–26035, 2025. 2

    work page 2025

  35. [35]

    One-step effective diffusion network for real-world image super-resolution.Advances in Neural Information Process- ing Systems, 37:92529–92553, 2024

    Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution.Advances in Neural Information Process- ing Systems, 37:92529–92553, 2024. 2

    work page 2024

  36. [36]

    Understanding and improving layer normaliza- tion.Advances in neural information processing systems, 32,

    Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normaliza- tion.Advances in neural information processing systems, 32,

    work page

  37. [37]

    One-step diffusion with distribution matching distillation

    Tianwei Yin, Micha ¨el Gharbi, Richard Zhang, Eli Shecht- man, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 6613–6623, 2024. 2

    work page 2024

  38. [38]

    Renderformer: Transformer-based neural rendering of triangle meshes with global illumination

    Chong Zeng, Yue Dong, Pieter Peers, Hongzhi Wu, and Xin Tong. Renderformer: Transformer-based neural rendering of triangle meshes with global illumination. InACM SIG- GRAPH 2025 Conference Papers, 2025. 2

    work page 2025

  39. [39]

    Rgbx: Image decomposition and synthesis us- ing material- and lighting-aware diffusion models

    Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloˇs Haˇsan. Rgbx: Image decomposition and synthesis us- ing material- and lighting-aware diffusion models. InACM SIGGRAPH 2024 Conference Papers, New York, NY , USA,

    work page 2024

  40. [40]

    Association for Computing Machinery. 2, 6, 7

    work page

  41. [41]

    Instantrestore: Single-step personalized face restoration with shared-image attention,

    Howard Zhang, Yuval Alaluf, Sizhuo Ma, Achuta Kadambi, Jian Wang, and Kfir Aberman. Instantrestore: Single-step personalized face restoration with shared-image attention,

    work page

  42. [42]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InCVPR, 2018. 5, 6 10 RenderFlow: Single-Step Neural Rendering via Flow Matching Supplementary Material A. Implementation Details A.1. Forward Rendering Our model is fine-tuned from the pre-trained Wan2.1 (1...

    work page 2018