ART-MPG: Autoregressive R/T Manifold Path Guiding
Summary
ART-MPG 提出了一种紧凑的神经自回归模型,用于流形路径引导(Manifold Path Guiding),通过拓扑条件化的 von Mises–Fisher 混合模型采样反射/透射链,在相同采样数下实现了比现有在线和离线渲染算法更低的均方误差。
View Cached Full Text
Cached at: 10/22/26, 04:18 PM
# ART-MPG: Autoregressive R/T Manifold Path Guiding
Source: [https://arxiv.org/html/2610.29049](https://arxiv.org/html/2610.29049)
Zehao ZhaoAffiliation:Ming Hsieh Department of Electrical and Computer Engineering, University of Southern CaliforniaXiaole ZhangAffiliation:Ming Hsieh Department of Electrical and Computer Engineering, University of Southern CaliforniaXiyu Deng††thanks:Corresponding author:xiyud@andrew\.cmu\.edu\.Affiliation:Department of Electrical and Computer Engineering, Carnegie Mellon University
###### Abstract
Specular chains concentrate important transport in small regions of path space and the structure depends on chain topology\. We introduceART\-MPG, a compact neural model that samples chain length, an autoregressive reflection/transmission \(R/T\) string, and a direction from a topology\-conditioned four\-component von Mises–Fisher \(vMF\) mixture\. We extend the repeated trial weighting in Manifold Path Guiding \(MPG\) by conditioning retry sampling on both chain length and the sampled R/T string, accounting for its known mass outside the stochastic re\-hit trials\. For a fixed proposal and target chain, this preserves the ideal expectation while reducing the expected retry count and reciprocal\-estimator variance compared with retries conditioned on chain length alone\. We distinguish the formal unbiasedness result from the capped and clamped renderer used in experiments\. We evaluate fourteen scenes across a range of samples per pixel \(SPP\) against existing online and offline rendering algorithms\. Our method consistently produces lower mean squared error at equal SPP than the online and offline baselines, with lower standard deviation across runs\.
## 1Introduction
Specular transport produces visually important paths that occupy narrow, structured regions of path space\. Ordinary path tracing can miss these paths, while samples that find them may carry substantial energy\. Manifold methods address the geometric difficulty by constructing an approximate seed chain and projecting it onto the manifold defined by reflection and refraction constraints\([Jakob and Marschner, 2012](https://arxiv.org/html/2610.29049#bib.bib3);[Hanika et al\., 2015](https://arxiv.org/html/2610.29049#bib.bib15);[Zeltner et al\., 2020](https://arxiv.org/html/2610.29049#bib.bib4)\)\.
Manifold Path Guiding \(MPG\) learns from previously found admissible chains\. It factorizes the seed proposal into a distribution over chain lengths, a conditional distribution over complete reflection/transmission \(R/T\) strings, and a type\-conditioned directional von Mises–Fisher \(vMF\) mixture, reconstructing these distributions online from local path records\([Fan et al\., 2023](https://arxiv.org/html/2610.29049#bib.bib5)\)\. This is a powerful representation, but its complete\-type tables grow with the number of R/T strings, its directional mixture grows with local records, and its repeated\-trial estimator factors out only the known length probability\.
While diffusion has long been used to generate 3D space representations\([Zhang et al\., 2026](https://arxiv.org/html/2610.29049#bib.bib23)\), its application in the light rendering problem is still understudied\. We introduceART\-MPG\(autoregressive R/T manifold path guiding\), a compact neural amortization of this MPG pipeline\. A prefix model assigns an exactly evaluable probability mass to each ordered R/T string, and a fixed four\-component vMF head predicts the direction density conditioned on the completed descriptor\. The exact type mass can then be divided out before re\-hit estimation\. Under a common mixed proposal and a fixed target chain, conditioning on both length and type, rather than length alone, preserves the ideal expectation while reducing the variance of the reciprocal estimator\. Under the same finite retry cap, it also cannot increase the timeout probability\. Our theoretical contribution is the comparison between conditioning retries on chain length alone and conditioning them on both chain length and the R/T string, rather than the established mixed factorization, vMF family, or geometric re\-hit identity\.
The many\-to\-one nature of manifold projection remains central: a pointwise directional density is not the probability of producing a projected chain\. We therefore separate three objects throughout the paper: the normalized neural seed proposal, the ideal descriptor\-conditioned chain estimator inherited from the Specular Manifold Sampling \(SMS\)/MPG re\-hit lineage, and the finite\-work capped/clamped renderer actually evaluated\. This distinction makes explicit how the theoretical guarantees relate to the evaluated implementation\.
Our contributions are:
1. 1\.We introduce a mathematical formulation for modeling reflection/transmission sequences in specular transport, coupled with a directional proposal conditioned on the sampled chain length and complete R/T sequence \(Section[5\.1](https://arxiv.org/html/2610.29049#S5.SS1),[5\.2](https://arxiv.org/html/2610.29049#S5.SS2)\)
2. 2\.We develop a descriptor\-conditioned re\-hit estimator that factors out the exactly evaluable length\-and\-type probability\. For a common proposal and a fixed target chain, we prove that this construction preserves the ideal expectation and reduces reciprocal\-estimator variance whenever the target type probability is less than one, with equality otherwise \(Section[5\.3](https://arxiv.org/html/2610.29049#S5.SS3), Equation[3](https://arxiv.org/html/2610.29049#S5.E3), and Proposition[1](https://arxiv.org/html/2610.29049#Thmproposition1)\)\.
## 2Related Work
#### Path guiding and learned proposals\.
Classical path guiding fits spatially varying directional distributions from transport samples, including parametric mixtures, product\-guided densities, and parallax\- or variance\-aware models\([Vorba et al\., 2014](https://arxiv.org/html/2610.29049#bib.bib10);[Herholz et al\., 2016](https://arxiv.org/html/2610.29049#bib.bib11);[Müller et al\., 2017](https://arxiv.org/html/2610.29049#bib.bib7);[Ruppert et al\., 2020](https://arxiv.org/html/2610.29049#bib.bib13);[Rath et al\., 2020](https://arxiv.org/html/2610.29049#bib.bib14)\)\. Neural approaches instead amortize sampleable densities with compact networks or invertible mappings\([Müller et al\., 2019](https://arxiv.org/html/2610.29049#bib.bib12);[Zheng and Zwicker, 2019](https://arxiv.org/html/2610.29049#bib.bib16);[Bako et al\., 2019](https://arxiv.org/html/2610.29049#bib.bib17);[Dong et al\., 2023](https://arxiv.org/html/2610.29049#bib.bib6)\)\. These works motivate our normalized neural proposal, but ordinary directional guiding does not by itself account for discrete specular topology or the many\-to\-one manifold projection\.
#### Specular transport\.
Manifold exploration, MNEE, and SMS construct paths by solving reflection and refraction constraints\([Jakob and Marschner, 2012](https://arxiv.org/html/2610.29049#bib.bib3);[Hanika et al\., 2015](https://arxiv.org/html/2610.29049#bib.bib15);[Zeltner et al\., 2020](https://arxiv.org/html/2610.29049#bib.bib4)\)\. SMS also supplies the repeated\-reconstruction lineage behind geometric reciprocal estimation\. Deterministic root\-finding methods, including Specular Polynomials, offer a complementary route to discovering admissible chains\([Fan et al\., 2024](https://arxiv.org/html/2610.29049#bib.bib18)\); our work addresses proposal learning for the manifold\-walk pipeline rather than replacing its solver\.
#### Relation to MPG\.
MPG already decomposes a seed law into length, complete R/T type, and a type\-conditioned directional vMF mixture\. It reconstructs these factors online from local path records and, in its Eq\. 15, holds length fixed during SMS\-style re\-hit trials\([Fan et al\., 2023](https://arxiv.org/html/2610.29049#bib.bib5)\)\.ART\-MPGdoes not claim the mixed factorization, vMF family, or generic re\-hit identity as new\. It neurally amortizes MPG with a prefix\-autoregressive type model and a fixed\-size four\-component vMF head, then extends the exact discrete factor in the retry estimator from length to length plus ordered type\. The latter change removes the type probability from stochastic re\-hit trials, as quantified in Proposition[1](https://arxiv.org/html/2610.29049#Thmproposition1)\.
## 3Background and Preliminaries
#### Monte Carlo light transport\.
For a surface pointy∈ℳ⊂ℝ3y\\in\\mathcal\{M\}\\subset\\mathbb\{R\}^\{3\}and a unit outgoing directionωo∈𝕊2\\omega\_\{o\}\\in\\mathbb\{S\}^\{2\}, the rendering equation writes outgoing radiance as\([Kajiya, 1986](https://arxiv.org/html/2610.29049#bib.bib1)\)
Lo\(y,ωo\)=Le\(y,ωo\)\+∫ℋy2fs\(y,ωi,ωo\)Li\(y,ωi\)\|cosθi\|dωi\.L\_\{o\}\(y,\\omega\_\{o\}\)=L\_\{e\}\(y,\\omega\_\{o\}\)\+\\int\_\{\\mathcal\{H\}\_\{y\}^\{2\}\}f\_\{s\}\(y,\\omega\_\{i\},\\omega\_\{o\}\)L\_\{i\}\(y,\\omega\_\{i\}\)\|\\cos\\theta\_\{i\}\|\\,\\mathrm\{d\}\\omega\_\{i\}\.Here,ℳ\\mathcal\{M\}denotes the scene surfaces,𝕊2=\{ω∈ℝ3:‖ω‖2=1\}\\mathbb\{S\}^\{2\}=\\\{\\omega\\in\\mathbb\{R\}^\{3\}:\\\|\\omega\\\|\_\{2\}=1\\\}is the unit sphere, andℋy2\\mathcal\{H\}\_\{y\}^\{2\}is the hemisphere above the surface atyy\. The termsLeL\_\{e\},LiL\_\{i\}, andfsf\_\{s\}denote emitted radiance, incident radiance, and the surface scattering function, respectively;θi\\theta\_\{i\}is the angle between the incident directionωi\\omega\_\{i\}and the surface normal\. Ifωi∼p\(⋅∣y,ωo\)\\omega\_\{i\}\\sim p\(\\cdot\\mid y,\\omega\_\{o\}\), a one\-sample estimator of the integral is the integrand divided byp\(ωi∣y,ωo\)p\(\\omega\_\{i\}\\mid y,\\omega\_\{o\}\), whereppis a density with respect to solid angle and is positive wherever the integrand is nonzero\. Consequently, a proposal that places probability near high\-contribution directions can reduce variance; one that undersamples a narrow contribution region produces rare, high\-weight samples\. Path guiding learns such proposals from transport observations\([Vorba et al\., 2014](https://arxiv.org/html/2610.29049#bib.bib10);[Müller et al\., 2017](https://arxiv.org/html/2610.29049#bib.bib7);[Ruppert et al\., 2020](https://arxiv.org/html/2610.29049#bib.bib13)\)\. In our setting, a seed direction initializes a nonlinear manifold solve, and multiple seeds can converge to the same final path\. The probability of reconstructing that path therefore depends on the total probability mass of those seeds, rather than the pointwise density of the particular seed that found it\.
#### Specular chains and separator vertices\.
Following manifold path sampling and MPG\([Jakob and Marschner, 2012](https://arxiv.org/html/2610.29049#bib.bib3);[Fan et al\., 2023](https://arxiv.org/html/2610.29049#bib.bib5)\), we decompose a light path at non\-specular*separator*vertices\. Between two consecutive separators, a specular chain contains ideal reflection or refraction vertices jointly constrained by geometry and Snell’s law\. We describe its discrete configuration and initial seed direction by
d=\(n,t1:n\),ti∈\{R,T\},ω∈𝕊2d=\(n,t\_\{1:n\}\),\\qquad t\_\{i\}\\in\\\{\\mathrm\{R\},\\mathrm\{T\}\\\},\\qquad\\omega\\in\\mathbb\{S\}^\{2\}wherennis the number of specular vertices,t1:nt\_\{1:n\}is their ordered R/T strings, andω\\omegais the initial seed direction\. Samplingnn,t1:nt\_\{1:n\}, andω\\omegaconstructs an initial path; a manifold walk then adjusts its vertices to satisfy the coupled constraints or reports failure\([Hanika et al\., 2015](https://arxiv.org/html/2610.29049#bib.bib15);[Zeltner et al\., 2020](https://arxiv.org/html/2610.29049#bib.bib4)\)\. The resulting chain contributes the usual product of endpoint factors, scattering terms, geometry terms, and transmittance\. The challenge is therefore not merely to predict a bright direction, but to allocate probability across disconnected length\-and\-type strata before predicting a direction within the selected stratum\.
#### Projection basins and chain probability\.
For fixed outer renderer statezz, letΦz\(d,ω\)\\Phi\_\{z\}\(d,\\omega\)denote the deterministic manifold solve\. The map from seeds to admissible chains is many\-to\-one: for a given descriptordd, seed directions within an angular*convergence basin*can project to the same canonical chain\. Ifq\(d,ω\)q\(d,\\omega\)is the mixed seed proposal, the probability of a chain is the integral ofqqover its preimage underΦz\\Phi\_\{z\}, not the pointwise density of the particular seed that found it\. SMS estimates the reciprocal basin mass by repeated reconstruction, while MPG factors known discrete probability mass from this trial process\([Zeltner et al\., 2020](https://arxiv.org/html/2610.29049#bib.bib4);[Fan et al\., 2023](https://arxiv.org/html/2610.29049#bib.bib5)\)\. Section[4](https://arxiv.org/html/2610.29049#S4)formalizes this chain\-space probability; Section[5\.3](https://arxiv.org/html/2610.29049#S5.SS3)describes the length\-and\-type stratification used byART\-MPG\.
#### Directional proposals on the sphere\.
For a unit mean directionμ∈𝕊2\\mu\\in\\mathbb\{S\}^\{2\}and concentrationκ≥0\\kappa\\geq 0, the vMF density on𝕊2\\mathbb\{S\}^\{2\}is
vMF\(ω,μ,κ\)=κ4πsinhκexp\(κμ⊤ω\),\\operatorname\{vMF\}\(\\omega;\\mu,\\kappa\)=\\frac\{\\kappa\}\{4\\pi\\sinh\\kappa\}\\exp\(\\kappa\\mu^\{\\top\}\\omega\),\(1\)with theκ→0\\kappa\\rightarrow 0limit equal to the uniform spherical density1/\(4π\)1/\(4\\pi\)\. A mixture∑jρjvMF\(ω,μj,κj\)\\sum\_\{j\}\\rho\_\{j\}\\operatorname\{vMF\}\(\\omega;\\mu\_\{j\},\\kappa\_\{j\}\)represents multiple directional modes while retaining a normalized, analytically evaluable, and directly sampleable density\. Neural path guiding, NPM, and MPG use vMF mixtures for spherical transport proposals\([Ruppert et al\., 2020](https://arxiv.org/html/2610.29049#bib.bib13);[Dong et al\., 2023](https://arxiv.org/html/2610.29049#bib.bib6);[Fan et al\., 2023](https://arxiv.org/html/2610.29049#bib.bib5)\)\.ART\-MPGpredicts a four\-component mixture conditioned on the complete sampled R/T descriptor\. Numerical conventions and exact sampling details are provided in Appendix[B\.2](https://arxiv.org/html/2610.29049#A2.SS2)\.
#### Discrete diffusion for R/T sequences\.
For a chain of lengthnn, the reflection/transmission sequencet1:n∈\{R,T\}nt\_\{1:n\}\\in\\\{\\mathrm\{R\},\\mathrm\{T\}\\\}^\{n\}specifies the interaction type at each specular vertex\. Discrete diffusion models such sequences through a forward corruption process and a learned reverse process\([Xu et al\., 2026](https://arxiv.org/html/2610.29049#bib.bib19);[Xu and Chi, 2024](https://arxiv.org/html/2610.29049#bib.bib20);[Banerjee et al\., 2025](https://arxiv.org/html/2610.29049#bib.bib21);[Hoshino et al\., 2026](https://arxiv.org/html/2610.29049#bib.bib22)\)\. The forward process progressively corrupts the categorical tokens, for example by masking or replacing them\. Conditioned on the corrupted sequence, the diffusion step, and the transport context, a denoising network predicts probabilities for the R/T tokens\. Generation proceeds through successive reverse steps to obtain a complete R/T sequence, which can then condition the directional proposal used to initialize a specular chain\.
## 4Problem Formulation
Letzzdenote the renderer state before seed sampling, including the separator vertices, the sampled emitter point, and the outcomes of preceding random choices\. The network takes the corresponding featuresx=ψ\(z\)x=\\psi\(z\)as input\. Our analysis assumes thatzzremains fixed during the initial solve and all subsequent trials\. Seed\-dependent sampling of glossy offsets falls outside these assumptions, as discussed in[SectionC\.1](https://arxiv.org/html/2610.29049#A3.SS1)\. Let𝒜\(z\)\\mathcal\{A\}\(z\)denote the set of admissible specular chains with at mostNmaxN\_\{\\max\}vertices, and letFz\(a\)F\_\{z\}\(a\)be the RGB contribution of chainaaunder statezz\. We seek to estimate
I\(z\)=∑a∈𝒜\(z\)Fz\(a\)\.I\(z\)=\\sum\_\{a\\in\\mathcal\{A\}\(z\)\}F\_\{z\}\(a\)\.We sample seeds from a learned distribution and obtain admissible chains through the manifold solver\. Each chain’s contribution is weighted by an estimate of its reciprocal sampling probability\. All identities and expectations below apply componentwise to the RGB channels\.
A seed consists of a descriptord=\(n,t1:n\)d=\(n,t\_\{1:n\}\)and an initial directionω∈𝕊2\\omega\\in\\mathbb\{S\}^\{2\}\. The seed space is
𝒴=⨆n=1Nmax⨆t1:n∈\{R,T\}n\(\{\(n,t1:n\)\}×𝕊2\)\.\\mathcal\{Y\}=\\bigsqcup\_\{n=1\}^\{N\_\{\\max\}\}\\bigsqcup\_\{t\_\{1:n\}\\in\\\{\\mathrm\{R\},\\mathrm\{T\}\\\}^\{n\}\}\\bigl\(\\\{\(n,t\_\{1:n\}\)\\\}\\times\\mathbb\{S\}^\{2\}\\bigr\)\.Conditioned onxx, we sample seeds from the normalized proposal
qθ\(d,ω∣x\)=qθ\(n∣x\)∏i=1nqθ\(ti∣t<i,n,x\)gθ\(ω∣d,x\)\.q\_\{\\theta\}\(d,\\omega\\mid x\)=q\_\{\\theta\}\(n\\mid x\)\\prod\_\{i=1\}^\{n\}q\_\{\\theta\}\(t\_\{i\}\\mid t\_\{<i\},n,x\)g\_\{\\theta\}\(\\omega\\mid d,x\)\.\(2\)This follows MPG’s length–type–direction factorization\. We parameterize the distributions with a neural network and model the R/T sequence autoregressively\.
LetΦz\(d,ω\)\\Phi\_\{z\}\(d,\\omega\)denote the output of the deterministic manifold solver\. For the estimator, we define
Φ¯z\(d,ω\)=\{a,Φz\(d,ω\)=aandd\(a\)=d,∅,otherwise\.\\bar\{\\Phi\}\_\{z\}\(d,\\omega\)=\\begin\{cases\}a,&\\Phi\_\{z\}\(d,\\omega\)=a\\ \\text\{and\}\\ d\(a\)=d,\\\\ \\varnothing,&\\text\{otherwise\}\.\\end\{cases\}The outcome∅\\varnothingcontributes zero to the estimator\. It includes both unsuccessful solves and solutions whose descriptors differ from the sampled descriptor\. Letν\\nube the product of counting measure over descriptors and surface measure on𝕊2\\mathbb\{S\}^\{2\}\. The probability of obtaining chainaais
Qθ\(a∣z\)=∫Φ¯z−1\(\{a\}\)qθ\(d,ω∣ψ\(z\)\)𝑑ν\(d,ω\)\.Q\_\{\\theta\}\(a\\mid z\)=\\int\_\{\\bar\{\\Phi\}\_\{z\}^\{\-1\}\(\\\{a\\\}\)\}q\_\{\\theta\}\(d,\\omega\\mid\\psi\(z\)\)\\,\\mathrm\{d\}\\nu\(d,\\omega\)\.The integral covers all seeds mapped toaabyΦ¯z\\bar\{\\Phi\}\_\{z\}\. Evaluating the proposal density at a single seed does not give this probability\.
We train the network offline using a fixed dataset of chains obtained by the manifold solver,
ℛ=\{\(xb,nb,tb,1:nb,ωb,eb\)\}b=1B,\\mathcal\{R\}=\\\{\(x\_\{b\},n\_\{b\},t\_\{b,1:n\_\{b\}\},\\omega\_\{b\},e\_\{b\}\)\\\}\_\{b=1\}^\{B\},whereωb\\omega\_\{b\}is the unit direction from the receiver to the first vertex of the converged chain, andeb≥0e\_\{b\}\\geq 0is a recorded scalar energy proxy\.[SectionB\.1](https://arxiv.org/html/2610.29049#A2.SS1)describes the data collection procedure and the definition ofebe\_\{b\}\. Neither the target directionωb\\omega\_\{b\}nor data from reference images are included inxbx\_\{b\}\. The network parameters remain fixed during rendering\. Data collection and training costs are excluded from the comparison at equal SPP\.
We assume that every contributing chain has positive sampling probability and that∑a∈𝒜\(z\)‖Fz\(a\)‖<∞\\sum\_\{a\\in\\mathcal\{A\}\(z\)\}\\\|F\_\{z\}\(a\)\\\|<\\infty\. The analysis uses independent trials under a fixed statezzand an exact test for whether two solutions represent the same chain\. Appendix[C\.1](https://arxiv.org/html/2610.29049#A3.SS1)gives the full assumptions and the proof of unbiasedness\. Section[5\.4](https://arxiv.org/html/2610.29049#S5.SS4)describes the implementation used in the experiments\.
## 5Algorithm Design
ART\-MPGfollows MPG’s factorization into chain length, R/T sequence, and direction, with two changes\. First, we replace MPG’s local tables of complete R/T sequences and directional kernels centered on nearby samples with a neural network of fixed size\. The network models the R/T sequence autoregressively and predicts a directional distribution conditioned on the chain length and complete sequence\. Second, MPG holds only the chain length fixed during repeated trials and resamples both the R/T sequence and direction\. We hold both the length and sequence fixed and resample only the direction\. The resulting estimate of the reciprocal conditional probability is divided by the exact joint probability of the length and sequence\.
### 5\.1Autoregressive descriptor and directional density
From Equation[2](https://arxiv.org/html/2610.29049#S4.E2), the probability of a descriptord=\(n,t1:n\)d=\(n,t\_\{1:n\}\)is
πθ\(d∣x\)=qθ\(n∣x\)∏i=1nqθ\(ti∣t<i,n,x\)\.\\pi\_\{\\theta\}\(d\\mid x\)=q\_\{\\theta\}\(n\\mid x\)\\prod\_\{i=1\}^\{n\}q\_\{\\theta\}\(t\_\{i\}\\mid t\_\{<i\},n,x\)\.The joint proposal can therefore be written asqθ\(d,ω∣x\)=πθ\(d∣x\)gθ\(ω∣d,x\)q\_\{\\theta\}\(d,\\omega\\mid x\)=\\pi\_\{\\theta\}\(d\\mid x\)g\_\{\\theta\}\(\\omega\\mid d,x\)\. Sampling starts with all descriptor tokens masked\. The network first predicts the distribution over chain lengths\. After samplingnn, it generates the R/T sequence in order, conditioning each prediction on the preceding symbols\. The descriptor probabilityπθ\(d∣x\)\\pi\_\{\\theta\}\(d\\mid x\)is evaluated exactly from the length probability and the conditional probabilities used during sampling\.
Once the descriptor is complete, the network predicts a mixture of the vMF densities defined in Equation[1](https://arxiv.org/html/2610.29049#S3.E1):
gθ\(ω∣d,x\)=∑j=1KρjvMF\(ω,μj,κj\)\.g\_\{\\theta\}\(\\omega\\mid d,x\)=\\sum\_\{j=1\}^\{K\}\\rho\_\{j\}\\operatorname\{vMF\}\(\\omega;\\mu\_\{j\},\\kappa\_\{j\}\)\.The number of componentsKKis fixed\. The weights satisfyρj≥0\\rho\_\{j\}\\geq 0and∑jρj=1\\sum\_\{j\}\\rho\_\{j\}=1, the mean directions satisfyμj∈𝕊2\\mu\_\{j\}\\in\\mathbb\{S\}^\{2\}, and the concentrations satisfyκj≥0\\kappa\_\{j\}\\geq 0\. All mixture parameters depend onddandxx\.
MPG also conditions its directional mixtures on the chain type, but constructs them from nearby records\. Our network predicts the mixture parameters directly from the descriptor and scene features\. Appendix[B\.2](https://arxiv.org/html/2610.29049#A2.SS2)gives the numerical details of density evaluation and sampling\. The network architecture and the encoding of features and descriptor tokens are described in Appendix[B\.1](https://arxiv.org/html/2610.29049#A2.SS1)\.
### 5\.2Offline objective
Training minimizes the negative log likelihood of the chain length, the R/T sequence, and the direction\. For the R/T sequence, we sample a prefix uniformly and multiply the loss for predicting the next symbol by the sequence length, giving an unbiased estimate of the loss over the complete sequence\. Each record is weighted using a logarithmic transformation of its recorded energy to moderate the influence of large values\. The training records are collected by the manifold integrator using online MPG, with each direction target pointing from the receiver to the first vertex of the converged chain\. Reference images are not used for training\. Appendix[B\.1](https://arxiv.org/html/2610.29049#A2.SS1)provides the full objective, the definition of the recorded energy, and the training procedure\. The trained model remains fixed during evaluation\.
### 5\.3Factoring type mass from re\-hit trials
For a chainaawith descriptord\(a\)d\(a\), the probability of obtainingaawhen its descriptor is held fixed is
Rθ\(a∣d\(a\),z\)=∫𝕊2𝟏\{Φ¯z\(d\(a\),ω\)=a\}gθ\(ω∣d\(a\),ψ\(z\)\)dω\.R\_\{\\theta\}\(a\\mid d\(a\),z\)=\\int\_\{\\mathbb\{S\}^\{2\}\}\\mathbf\{1\}\\\{\\bar\{\\Phi\}\_\{z\}\(d\(a\),\\omega\)=a\\\}g\_\{\\theta\}\(\\omega\\mid d\(a\),\\psi\(z\)\)\\,\\mathrm\{d\}\\omega\.Writingd\(a\)=\(n,t1:n\)d\(a\)=\(n,t\_\{1:n\}\)andx=ψ\(z\)x=\\psi\(z\), the chain probability isQθ\(a∣z\)=qNqTRθQ\_\{\\theta\}\(a\\mid z\)=q\_\{N\}q\_\{T\}R\_\{\\theta\}, where
qN=qθ\(n∣x\),qT=∏i=1nqθ\(ti∣t<i,n,x\)\.q\_\{N\}=q\_\{\\theta\}\(n\\mid x\),\\qquad q\_\{T\}=\\prod\_\{i=1\}^\{n\}q\_\{\\theta\}\(t\_\{i\}\\mid t\_\{<i\},n,x\)\.BothqNq\_\{N\}andqTq\_\{T\}can be evaluated exactly\. SMS estimates reciprocal probabilities by counting independent trials until the same chain is obtained again\. MPG evaluatesqNq\_\{N\}separately and holds the length fixed while resampling the R/T sequence and direction on each trial\([Fan et al\., 2023](https://arxiv.org/html/2610.29049#bib.bib5), Eq\. 15\)\. We also hold the R/T sequence fixed and resample only the direction\. The trial count estimates1Rθ\\frac\{1\}\{R\_\{\\theta\}\}\. Dividing this count byqNqTq\_\{N\}q\_\{T\}gives an estimate of1Qθ\(a∣z\)\\frac\{1\}\{Q\_\{\\theta\}\(a\\mid z\)\}\.
###### Proposition 1\(Effect of fixing the R/T sequence\)\.
Fix a chainaaand statezz, and use the same proposal and solver in both constructions\. LetqT∈\(0,1\]q\_\{T\}\\in\(0,1\]be the conditional probability of its R/T sequence and letR=Rθ\(a∣d\(a\),z\)∈\(0,1\]R=R\_\{\\theta\}\(a\\mid d\(a\),z\)\\in\(0,1\]\. With independent trials and exact comparison of chain solutions, the number of trials required to obtainaasatisfiesJn∼Geom\(qTR\)J\_\{n\}\\sim\\operatorname\{Geom\}\(q\_\{T\}R\)andJd∼Geom\(R\),J\_\{d\}\\sim\\operatorname\{Geom\}\(R\),when only the length is fixed and when the complete descriptor is fixed, respectively\. Both counts include the successful trial\. Then
𝔼\[Jd\]=1/R≤1/\(qTR\)=𝔼\[Jn\],\\displaystyle\\mathbb\{E\}\[J\_\{d\}\]=1/R\\leq 1/\(q\_\{T\}R\)=\\mathbb\{E\}\[J\_\{n\}\],Var\(Jn\)−Var\(Jd/qT\)=1−qTqT2R≥0\.\\displaystyle\\operatorname\{Var\}\(J\_\{n\}\)\-\\operatorname\{Var\}\(J\_\{d\}/q\_\{T\}\)=\\frac\{1\-q\_\{T\}\}\{q\_\{T\}^\{2\}R\}\\geq 0\.For any nonnegative integer trial limitMM,
Pr\[Jd\>M\]=\(1−R\)M≤\(1−qTR\)M=Pr\[Jn\>M\]\.\\Pr\[J\_\{d\}\>M\]=\(1\-R\)^\{M\}\\leq\(1\-q\_\{T\}R\)^\{M\}=\\Pr\[J\_\{n\}\>M\]\.
These results follow from𝔼\[J\]=1/p\\mathbb\{E\}\[J\]=1/p,Var\(J\)=\(1−p\)/p2\\operatorname\{Var\}\(J\)=\(1\-p\)/p^\{2\}, andPr\[J\>M\]=\(1−p\)M\\Pr\[J\>M\]=\(1\-p\)^\{M\}forJ∼Geom\(p\)J\\sim\\operatorname\{Geom\}\(p\)\. To estimateI\(z\)I\(z\), drawD∼πθ\(⋅∣x\)D\\sim\\pi\_\{\\theta\}\(\\cdot\\mid x\)andΩ0∼gθ\(⋅∣D,x\)\\Omega\_\{0\}\\sim g\_\{\\theta\}\(\\cdot\\mid D,x\), and setA=Φ¯z\(D,Ω0\)A=\\bar\{\\Phi\}\_\{z\}\(D,\\Omega\_\{0\}\)\. IfA≠∅A\\neq\\varnothing, keepDDandzzfixed and draw new directions independently until the solver returns the same chainAA\. LetJdJ\_\{d\}count these additional trials, including the successful trial\. The estimator is
I^ART\-MPG=\{0,A=∅,Fz\(A\)Jd/\(qNqT\),A≠∅\.\\widehat\{I\}\_\{\\mathrm\{ART\\text\{\-\}MPG\}\}=\\begin\{cases\}0,&A=\\varnothing,\\\\ F\_\{z\}\(A\)J\_\{d\}/\(q\_\{N\}q\_\{T\}\),&A\\neq\\varnothing\.\\end\{cases\}\(3\)Under the assumptions stated in Appendix[C\.1](https://arxiv.org/html/2610.29049#A3.SS1), this estimator is unbiased forI\(z\)I\(z\)\. The proof is given in Appendix[C\.1](https://arxiv.org/html/2610.29049#A3.SS1)and follows the repeated trial argument of SMS and MPG\([Zeltner et al\., 2020](https://arxiv.org/html/2610.29049#bib.bib4);[Fan et al\., 2023](https://arxiv.org/html/2610.29049#bib.bib5)\)\.
### 5\.4Finite\-work renderer
The renderer starts from statez0z\_\{0\}, samples a descriptorD=\(N,T1:N\)D=\(N,T\_\{1:N\}\), and draws a direction from the corresponding vMF mixture\. The manifold solver then produces a chainAA, subject to the descriptor condition defined above\. The initial solve also determines a statezηz\_\{\\eta\}containing the sampled offset normals\. Bothzηz\_\{\\eta\}andDDremain fixed during subsequent trials, which sample new directions independently\. Two solutions are treated as the same chain when their shape sequences and R/T sequences agree and their first segment directions agree within a numerical tolerance\. LetJimplJ\_\{\\rm impl\}denote the number of trials required to obtain a match without a trial limit, including the successful trial\. We setJimpl=∞J\_\{\\rm impl\}=\\inftyif the initial solve fails\. This numerical comparison differs from the exact equality assumed in the definition ofJdJ\_\{d\}\.
Before applying the weights for chain sampling, the contribution isHzη\(A\)=frecv\(zη,A\)Gzη\(A\)poff\(zη,A\),H\_\{z\_\{\\eta\}\}\(A\)=\\frac\{f\_\{\\rm recv\}\(z\_\{\\eta\},A\)G\_\{z\_\{\\eta\}\}\(A\)\}\{p\_\{\\rm off\}\(z\_\{\\eta\},A\)\},wherefrecvf\_\{\\rm recv\}is the receiver BSDF factor,Gzη\(A\)G\_\{z\_\{\\eta\}\}\(A\)is the remaining chain contribution, andpoffp\_\{\\rm off\}is the density of the sampled offset normals\. Samples with invalid offset densities contribute zero\. For finite nonnegative contributions, the implementation computes
I^fw=\{clipCfpeak\(Hzη\(A\)min\{Cw,JimplqT\}\)max\{εN,qN\},A≠∅,Jimpl≤M,0,otherwise\.\\widehat\{I\}\_\{\\mathrm\{fw\}\}=\\begin\{cases\}\\displaystyle\\frac\{\\operatorname\{clip\}^\{\\mathrm\{peak\}\}\_\{C\_\{f\}\}\\left\(H\_\{z\_\{\\eta\}\}\(A\)\\min\\left\\\{C\_\{w\},\\frac\{J\_\{\\rm impl\}\}\{q\_\{T\}\}\\right\\\}\\right\)\}\{\\max\\\{\\varepsilon\_\{N\},q\_\{N\}\\\}\},&A\\neq\\varnothing,\\ J\_\{\\rm impl\}\\leq M,\\\\\[4\.30554pt\] 0,&\\text\{otherwise\}\.\\end\{cases\}\(4\)Here,MMlimits the number of trials,CwC\_\{w\}boundsJimpl/qTJ\_\{\\rm impl\}/q\_\{T\}, andεN\\varepsilon\_\{N\}is a positive lower bound applied to the length probability\. The operatorclipCfpeak\\operatorname\{clip\}^\{\\mathrm\{peak\}\}\_\{C\_\{f\}\}rescales all RGB channels by the same factor when the largest channel exceedsCfC\_\{f\}\. This rescaling precedes division bymax\{εN,qN\}\\max\\\{\\varepsilon\_\{N\},q\_\{N\}\\\}\. The pseudocode below gives the sampling and weighting procedure\. Appendix[C\.2](https://arxiv.org/html/2610.29049#A3.SS2)analyzes the effect of limiting the number of trials\.
At an eligible smooth receiver, the renderer evaluates the manifold contribution alongside ordinary next event estimation and continues the path through BSDF sampling\. The emitter and BSDF strategies are combined using MIS with the power heuristic\([Veach and Guibas, 1995](https://arxiv.org/html/2610.29049#bib.bib2)\)\. The manifold contribution is added separately and is not included in these MIS weights\. Its proposal does not use a defensive mixture\. When the routing option is enabled and manifold sampling has been invoked, later contributions from finite emitters reached after a caustic caster are suppressed\. Contributions from the environment remain enabled\. The proof in Appendix[C\.1](https://arxiv.org/html/2610.29049#A3.SS1)establishes unbiasedness of the conditional chain estimator under the stated assumptions\. It does not establish unbiasedness of the complete renderer used in the experiments\.
Algorithm 1Finite\-workART\-MPGsample used in the experiments1:Pre\-seed state
z0z\_\{0\}, features
x=ψ\(z0\)x=\\psi\(z\_\{0\}\);
M,Cw,CfM,C\_\{w\},C\_\{f\}\.
2:Sample
D=\(N,T1:N\)∼πθ\(⋅∣x\)D=\(N,T\_\{1:N\}\)\\sim\\pi\_\{\\theta\}\(\\cdot\\mid x\); retain
qN,qTq\_\{N\},q\_\{T\}\.
3:Draw
Ω0∼gθ\(⋅∣D,x\)\\Omega\_\{0\}\\sim g\_\{\\theta\}\(\\cdot\\mid D,x\); solve for
A,poff,zηA,p\_\{\\rm off\},z\_\{\\eta\}\.
4:if
A=∅A=\\varnothingor
poffp\_\{\\rm off\}is nonfinite or
poff≤10−12p\_\{\\rm off\}\\leq 10^\{\-12\}then
5:return
00
6:endif
7:for
j=1,…,Mj=1,\\ldots,Mdo
8:Draw
Ωj∼gθ\(⋅∣D,x\)\\Omega\_\{j\}\\sim g\_\{\\theta\}\(\\cdot\\mid D,x\)and solve with frozen
zηz\_\{\\eta\}\.
9:ifshape sequence, type code, and first\-direction root match
AAthen
10:
Jimpl←jJ\_\{\\rm impl\}\\leftarrow j;break
11:endif
12:endfor
13:ifno match occurredthen
14:return
00
15:endif
16:
V←Hzη\(A\)min\{Cw,Jimpl/qT\}V\\leftarrow H\_\{z\_\{\\eta\}\}\(A\)\\min\\\{C\_\{w\},J\_\{\\rm impl\}/q\_\{T\}\\\}\.
17:if
VVis nonfinite or
maxcVc<0\\max\_\{c\}V\_\{c\}<0then
18:return
00
19:endif
20:Rescale all RGB channels together if
maxcVc\>Cf\\max\_\{c\}V\_\{c\}\>C\_\{f\}\.
21:return
V/max\{10−6,qN\}V/\\max\\\{10^\{\-6\},q\_\{N\}\\\}\.
## 6Experiments, Results and Discussion
#### Experimental setup\.
We evaluate fourteen64×6464\\times 64scene configurations at\{12,24,36,48,60\}\\\{12,24,36,48,60\\\}samples per pixel \(SPP\)\. Every method is compared with the same fixed 32,768\-SPP numerical reference for each scene\. For every scene–method–SPP cell, we use the same ten fixed runs, denoted D01–D10, yielding a complete14×6×5×10=4,20014\\times 6\\times 5\\times 10=4\{,\}200\-image result matrix\.
We measure full\-frame mean squared error \(MSE\) on raw linear high\-dynamic\-range \(HDR\) RGB values before tone mapping\. Four\-channel OpenEXR arrays loaded by OpenCV use blue–green–red–alpha \(BGRA\) channel order\. We discard alpha, reorder the remaining blue–green–red channels to RGB, and do not premultiply color by alpha\. Images containing nonfinite values are rejected\.
Curves report the median across independent experiment runs D01–D10\. The 60\-SPP annotations additionally report the sample standard deviation of those ten values using denominatorn−1n\-1, while Appendix Table[A1](https://arxiv.org/html/2610.29049#A1.T1)also reports the arithmetic mean\. These statistics describe the ten fixed runs and are not confidence intervals\.
### 6\.1Comparisons
The benchmark comprises 14 scenes,Living,Lamp,Plane,Slabs,Stone,Pool,Sphere,Water,Ring\-and\-Slab,Pillars,Slab\-×1\\times 1,Slab\-×3\\times 3,Pillar\-×3\\times 3, andCornell Glass\. We compareART\-MPGwith a direction\-only adaptation of Neural Parametric Mixtures \(NPM\)\([Dong et al\., 2023](https://arxiv.org/html/2610.29049#bib.bib6)\); our implementation of Manifold Path Guiding \(MPG\)\([Fan et al\., 2023](https://arxiv.org/html/2610.29049#bib.bib5)\); a fixed\-chain diagnostic based on Specular Manifold Sampling \(SMS\)\([Zeltner et al\., 2020](https://arxiv.org/html/2610.29049#bib.bib4)\); Practical Path Guiding \(PPG\)\([Müller et al\., 2017](https://arxiv.org/html/2610.29049#bib.bib7)\); and unidirectional path tracing \(PT\)\([Kajiya, 1986](https://arxiv.org/html/2610.29049#bib.bib1)\)\.




Figure 1:Equal\-SPP results forLiving,Water,Stone, andPool, from top to bottom\. Curves show median full\-frame MSE across ten fixed runs; 60\-SPP annotations additionally show sample standard deviation\. Image crops come from one fixed 60\-SPP visualization run, whereas quantitative values use separate64×6464\\times 64full\-frame renders\. The displayed crops are not median images\.Figure[1](https://arxiv.org/html/2610.29049#S6.F1)shows four representative scenes\. Each strip combines a reference overview and visual\-only D01 crops from the native1152×11521152\\times 115260\-SPP film with median curves computed from the separate64×6464\\times 64quantitative renders\. The displayed crop illustrates one fixed run and is not a median image\. The remaining ten scene panels appear in Appendix[A\.4](https://arxiv.org/html/2610.29049#A1.SS4)\.
### 6\.2Evaluation
Table 1:Median full\-frame linear\-RGB MSE across D01–D10 for each scene at 60 SPP and64×6464\\times 64resolution\. Lower is better\.At 60 SPP,ART\-MPGhas the lowest median MSE on13/14scenes; SMS is lower onCornell Glass\. Across all five budgets,ART\-MPGhas the lowest median in62/70scene–SPP cells, and its median curve is non\-increasing at every adjacent budget on12/14scenes\. It has the lowest median at all five budgets on 11/14 scenes; nine scenes satisfy both this condition and monotonicity\. Appendix[A\.2](https://arxiv.org/html/2610.29049#A1.SS2)gives the complete rank and pairwise\-win counts, while Appendix Table[A1](https://arxiv.org/html/2610.29049#A1.T1)gives every 60\-SPP median, mean, and sample standard deviation\.
The only median upturns forART\-MPGoccur onLamp, from 0\.001974 at 24 SPP to 0\.001997 at 36 SPP, and onPlane, from 0\.2749 to 0\.2758 between 12 and 24 SPP and from 0\.2725 to 0\.2728 between 48 and 60 SPP\. These small increases are retained in the appendix panels\.
### 6\.3Discussion
The evaluation supports a focused conclusion: on this controlled benchmark,ART\-MPGusually gives lower linear\-HDR error per scheduled camera sample than the five comparison methods\. The result is strongest at 60 SPP and is broadly consistent across budgets, whileCornell Glassand the smallLamp/Planeupturns identify the observed exceptions\.
The methods operate under different learning regimes\.ART\-MPGand NPM are trained offline and frozen before rendering, MPG adapts online, and PPG updates its guiding state during rendering; PT and the fixed\-chain SMS diagnostic use fixed configurations\. The experiment therefore controls scheduled camera samples\.
NPM accesses trainable spatial grids sparsely, whereasART\-MPGevaluates a dense multilayer perceptron\. Execution details appear in Appendix[B\.3](https://arxiv.org/html/2610.29049#A2.SS3)\.
## 7Scope and Limitations
Our evidence supports a focused descriptive claim: over the ten runs on this controlled 14\-scene benchmark,ART\-MPGusually has lower median linear\-HDR error per scheduled camera sample than the five comparators\. Equal SPP does not imply equal wall\-clock cost; the neural guide is fitted offline per scene, its replay/training cost is excluded\. MPG performs online fitting inside its recorded schedule\. These are distinct operating regimes rather than a matched\-time comparison\.
Quantitative results use64×6464\\times 64images and one fixed32,76832\{,\}768\-SPP reference seed\. Several scenes share base geometry, and some benchmark families informed development, limiting claims of scene\-level independence or out\-of\-distribution generalization\. These boundaries motivate further research on matched\-time, reference\-convergence, clamp, and topology research for different SPP observations\.
## 8Conclusion
ART\-MPGmodels the complete low\-dimensional seed of a specular chain: its length, ordered reflection/transmission topology, and topology\-conditioned direction\. The controlled evaluation comparesART\-MPG, NPM, MPG, fixed\-chain SMS, PPG, and PT on 14 scenes at\{12,24,36,48,60\}\\\{12,24,36,48,60\\\}SPP\. Across 10 random seeds,ART\-MPGhas the lowest median MSE in 62/70 scene–SPP cells and in 13/14 scenes at 60 SPP\. It records 335/350 strict pairwise median wins over the five comparators across all budgets and 68/70 at 60 SPP; its median curve is non\-increasing on 12/14 scenes\. The only median upturns are small changes onLampandPlane, andCornell Glassis the sole 60\-SPP rank exception\. The results establish a sample\-budget comparison, not a matched\-time comparison, and supportART\-MPGas a compact proposal for structured\-caustic transport\.
## Reproducibility statement
The supplementary artifacts record the scenes, seeds, model exports, renderer settings, reference hashes, and per\-image metric rows\. Together with Section[6](https://arxiv.org/html/2610.29049#S6)and the complete pseudocode and proof in Appendix[C\.1](https://arxiv.org/html/2610.29049#A3.SS1), they distinguish the ideal estimator from the finite\-work profile evaluated in the paper\. The artifact contains the complete14×6×5×10=4,20014\\times 6\\times 5\\times 10=4\{,\}200rows for ART\-MPG, NPM, MPG, fixed\-chain SMS, PPG, and PT at\{12,24,36,48,60\}\\\{12,24,36,48,60\\\}SPP\. The reported curves are ordinary cellwise medians without dispersion bands\. Their 60\-SPP numeric annotations and the appendix table retain raw sample standard deviations with denominatorn−1n\-1\.
## AI use statement
Generative AI assisted with pseudocode, analysis/visualization code, and consistency checks, and performed the disclosed unblinded visual QA\. The authors verified AI\-assisted content against source code, execution receipts, tests, and rendered artifacts; AI was not used to synthesize measured values or replace renderer outputs\. The authors take responsibility for the final manuscript and artifacts\.
## References
- Bakoet al\.\(2019\)S\. Bako, M\. Meyer, T\. DeRose, and P\. SenOffline deep importance sampling for monte carlo path tracing\.Computer Graphics Forum38\(7\)\.Cited by:[§2](https://arxiv.org/html/2610.29049#S2.SS0.SSS0.Px1.p1.1)\.
- Banerjeeet al\.\(2025\)A\. Banerjee, X\. Xu, C\. Moosmüller, and H\. LeeAdaptive multimodal protein plug\-and\-play with diffusion\-based priors\.In2025 IEEE 10th International Workshop on Computational Advances in Multi\-Sensor Adaptive Processing \(CAMSAP\),pp\. 226–230\.Cited by:[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px5.p1.1)\.
- Donget al\.\(2023\)H\. Dong, G\. Wang, and S\. LiNeural parametric mixtures for path guiding\.InACM SIGGRAPH Conference Proceedings,pp\. 1–10\.External Links:[Document](https://dx.doi.org/10.1145/3588432.3591533)Cited by:[§2](https://arxiv.org/html/2610.29049#S2.SS0.SSS0.Px1.p1.1),[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px4.p1.2),[§6\.1](https://arxiv.org/html/2610.29049#S6.SS1.p1.1)\.
- Fanet al\.\(2024\)Z\. Fan, J\. Guo, Y\. Wang, T\. Xiao, H\. Zhang, C\. Zhou, Z\. Chen, P\. Hong, Y\. Guo, and L\. YanSpecular polynomials\.ACM Transactions on Graphics43\(4\),pp\. 1–13\.External Links:[Document](https://dx.doi.org/10.1145/3658132)Cited by:[§2](https://arxiv.org/html/2610.29049#S2.SS0.SSS0.Px2.p1.1)\.
- Fanet al\.\(2023\)Z\. Fan, P\. Hong, J\. Guo, C\. Zou, Y\. Guo, and L\. YanManifold path guiding for importance sampling specular chains\.ACM Transactions on Graphics42\(6\)\.External Links:[Document](https://dx.doi.org/10.1145/3618360)Cited by:[§C\.1](https://arxiv.org/html/2610.29049#A3.SS1.p2.1),[§1](https://arxiv.org/html/2610.29049#S1.p2.1),[§2](https://arxiv.org/html/2610.29049#S2.SS0.SSS0.Px3.p1.1),[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px2.p1.1),[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px3.p1.1),[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px4.p1.2),[§5\.3](https://arxiv.org/html/2610.29049#S5.SS3.p1.3),[§5\.3](https://arxiv.org/html/2610.29049#S5.SS3.p2.2),[§6\.1](https://arxiv.org/html/2610.29049#S6.SS1.p1.1)\.
- Hanikaet al\.\(2015\)J\. Hanika, M\. Droske, and L\. FascioneManifold next event estimation\.Computer Graphics Forum34\(4\)\.External Links:[Document](https://dx.doi.org/10.1111/cgf.12681)Cited by:[§1](https://arxiv.org/html/2610.29049#S1.p1.1),[§2](https://arxiv.org/html/2610.29049#S2.SS0.SSS0.Px2.p1.1),[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px2.p1.2)\.
- Herholzet al\.\(2016\)S\. Herholz, O\. Elek, J\. Vorba, H\. P\. A\. Lensch, and J\. KřivánekProduct importance sampling for light transport path guiding\.Computer Graphics Forum35\(4\)\.Cited by:[§2](https://arxiv.org/html/2610.29049#S2.SS0.SSS0.Px1.p1.1)\.
- Hoshinoet al\.\(2026\)K\. Hoshino, Y\. S\. Aluru, X\. Deng, and Y\. NakahiraTraining\-free guided diffusion for planning: a unified framework via doob’s h\-transform with safety guarantees\.InForty\-third International Conference on Machine Learning,Cited by:[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px5.p1.1)\.
- Jakob and Marschner \(2012\)W\. Jakob and S\. MarschnerManifold exploration: a markov chain monte carlo technique for rendering scenes with difficult specular transport\.ACM Transactions on Graphics31\(4\)\.External Links:[Document](https://dx.doi.org/10.1145/2185520.2185554)Cited by:[§1](https://arxiv.org/html/2610.29049#S1.p1.1),[§2](https://arxiv.org/html/2610.29049#S2.SS0.SSS0.Px2.p1.1),[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px2.p1.1)\.
- Kajiya \(1986\)J\. T\. KajiyaThe rendering equation\.InProceedings of SIGGRAPH,pp\. 143–150\.External Links:[Document](https://dx.doi.org/10.1145/15922.15902)Cited by:[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px1.p1.1),[§6\.1](https://arxiv.org/html/2610.29049#S6.SS1.p1.1)\.
- Mardia and Jupp \(2000\)K\. V\. Mardia and P\. E\. JuppDirectional statistics\.Wiley\.Cited by:[§B\.2](https://arxiv.org/html/2610.29049#A2.SS2.p1.1)\.
- Mülleret al\.\(2017\)T\. Müller, M\. Gross, and J\. NovákPractical path guiding for efficient light\-transport simulation\.InComputer Graphics Forum,Vol\.36,pp\. 91–100\.Cited by:[§2](https://arxiv.org/html/2610.29049#S2.SS0.SSS0.Px1.p1.1),[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px1.p1.2),[§6\.1](https://arxiv.org/html/2610.29049#S6.SS1.p1.1)\.
- Mülleret al\.\(2019\)T\. Müller, B\. McWilliams, F\. Rousselle, M\. Gross, and J\. NovákNeural importance sampling\.ACM Transactions on Graphics38\(5\)\.Cited by:[§2](https://arxiv.org/html/2610.29049#S2.SS0.SSS0.Px1.p1.1)\.
- Rathet al\.\(2020\)A\. Rath, P\. Grittmann, S\. Herholz, P\. Vévoda, P\. Slusallek, and J\. KřivánekVariance\-aware path guiding\.ACM Transactions on Graphics39\(4\)\.Cited by:[§2](https://arxiv.org/html/2610.29049#S2.SS0.SSS0.Px1.p1.1)\.
- Ruppertet al\.\(2020\)L\. Ruppert, S\. Herholz, and H\. P\. A\. LenschRobust fitting of parallax\-aware mixtures for path guiding\.ACM Transactions on Graphics39\(4\)\.Cited by:[§2](https://arxiv.org/html/2610.29049#S2.SS0.SSS0.Px1.p1.1),[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px1.p1.2),[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px4.p1.2)\.
- Veach and Guibas \(1995\)E\. Veach and L\. J\. GuibasOptimally combining sampling techniques for monte carlo rendering\.InProceedings of SIGGRAPH,pp\. 419–428\.External Links:[Document](https://dx.doi.org/10.1145/218380.218498)Cited by:[§5\.4](https://arxiv.org/html/2610.29049#S5.SS4.p3.1)\.
- Vorbaet al\.\(2014\)J\. Vorba, O\. Karlík, M\. Šík, T\. Ritschel, and J\. KřivánekOn\-line learning of parametric mixture models for light transport simulation\.ACM Transactions on Graphics33\(4\)\.External Links:[Document](https://dx.doi.org/10.1145/2601097.2601203)Cited by:[§2](https://arxiv.org/html/2610.29049#S2.SS0.SSS0.Px1.p1.1),[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px1.p1.2)\.
- Wood \(1994\)A\. T\. A\. WoodSimulation of the von mises fisher distribution\.Communications in Statistics – Simulation and Computation23\(1\),pp\. 157–164\.External Links:[Document](https://dx.doi.org/10.1080/03610919408813161)Cited by:[§B\.2](https://arxiv.org/html/2610.29049#A2.SS2.p1.2)\.
- Xu and Chi \(2024\)X\. Xu and Y\. ChiProvably robust score\-based diffusion posterior sampling for plug\-and\-play image reconstruction\.Advances in Neural Information Processing Systems37,pp\. 36148–36184\.Cited by:[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px5.p1.1)\.
- Xuet al\.\(2026\)X\. Xu, Z\. Zhang, Y\. Nakahira, G\. Qu, and Y\. ChiPolynomial convergence of riemannian diffusion models\.arXiv preprint arXiv:2601\.02499\.Cited by:[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px5.p1.1)\.
- Yinet al\.\(2026\)S\. Yin, Z\. Zhao, K\. Asadi, R\. Liu, Y\. Lu, S\. Mei, H\. Cui, L\. Simon, Z\. Shi, and H\. FiroozRecursive self\-improvement via on\-policy distillation for reasoning\.arXiv preprint arXiv:2609\.30652\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2609.30652),[Link](https://arxiv.org/abs/2609.30652)Cited by:[Appendix D](https://arxiv.org/html/2610.29049#A4.SS0.SSS0.Px1.p1.1)\.
- Yuehet al\.\(2026\)C\. Yueh, A\. Zhao, A\. Nayyar, and R\. JainBayesian learning in episodic zero\-sum games\.arXiv preprint arXiv:2603\.20604\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2603.20604),[Link](https://arxiv.org/abs/2603.20604)Cited by:[Appendix D](https://arxiv.org/html/2610.29049#A4.SS0.SSS0.Px2.p1.1)\.
- Zeltneret al\.\(2020\)T\. Zeltner, I\. Georgiev, and W\. JakobSpecular manifold sampling for rendering high\-frequency caustics and glints\.ACM Transactions on Graphics39\(4\)\.Cited by:[§C\.1](https://arxiv.org/html/2610.29049#A3.SS1.p2.1),[§1](https://arxiv.org/html/2610.29049#S1.p1.1),[§2](https://arxiv.org/html/2610.29049#S2.SS0.SSS0.Px2.p1.1),[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px2.p1.2),[§3](https://arxiv.org/html/2610.29049#S3.SS0.SSS0.Px3.p1.1),[§5\.3](https://arxiv.org/html/2610.29049#S5.SS3.p2.2),[§6\.1](https://arxiv.org/html/2610.29049#S6.SS1.p1.1)\.
- Zhanget al\.\(2026\)Z\. Zhang, S\. Li, Y\. Nakahira, P\. Bogdan, and G\. QuHilbertGen\-3d:hilbert–multifractal conditioning for topology\-aware 3d generation\.External Links:2610\.22720,[Link](https://arxiv.org/abs/2610.22720)Cited by:[§1](https://arxiv.org/html/2610.29049#S1.p3.1)\.
- Zhaoet al\.\(2025\)Z\. Zhao, A\. Shukla, R\. Jain, and V\. SubramanianEfficient decentralized learning of generalized quantal response equilibrium\.arXiv preprint arXiv:2507\.09928\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2507.09928),[Link](https://arxiv.org/abs/2507.09928)Cited by:[Appendix D](https://arxiv.org/html/2610.29049#A4.SS0.SSS0.Px3.p1.1)\.
- Zheng and Zwicker \(2019\)Q\. Zheng and M\. ZwickerLearning to importance sample in primary sample space\.Computer Graphics Forum38\(2\),pp\. 169–179\.External Links:[Document](https://dx.doi.org/10.1111/cgf.13628)Cited by:[§2](https://arxiv.org/html/2610.29049#S2.SS0.SSS0.Px1.p1.1)\.
## Appendix AAdditional Results
### A\.1Evaluation grid
The reported evaluation comparesART\-MPG, NPM, MPG, fixed\-chain SMS, PPG, and PT at\{12,24,36,48,60\}\\\{12,24,36,48,60\\\}SPP on fourteen configurations:Living,Lamp,Plane,Slabs,Stone,Pool,Sphere,Water,Ring\-and\-Slab,Pillars,Slab\-×1\\times 1,Slab\-×3\\times 3,Pillar\-×3\\times 3, andCornell Glass\. Every quantitative cell is full\-frame MSE on raw linear\-HDR RGB at64×6464\\times 64resolution against the same fixed 32,768\-SPP numerical reference for that scene\.
The SMS curve is the hash\-bound fixed\-chain configuration used by this sweep: the establishedLamp/Slabspresets are retained, and the remaining scenes use the frozen descriptor recorded by the protocol\. We therefore interpret SMS as a fixed\-chain diagnostic, not as a variable\-chain SMS result\.
The ten fixed runs are denoted D01–D10\. Each run contains the complete14×6×514\\times 6\\times 5grid, yielding14×6×5×10=4,20014\\times 6\\times 5\\times 10=4\{,\}200measured images\. The same run IDs are used for all methods and scenes\.
### A\.2Median outcome
ART\-MPG has the lowest median MSE in 62/70 scene–SPP cells, including 13/14 scenes at 60 SPP\. Its median curve is non\-increasing on 12/14 scenes and strictly rank 1 at every displayed budget on 11/14 scenes; nine scenes satisfy both properties\. It records 335/350 strict pairwise median wins against the five comparators across all budgets and 68/70 at 60 SPP\.Cornell Glassis the sole 60\-SPP rank exception\.
The only ART\-MPG median upturns areLamp, from 0\.001974 at 24 SPP to 0\.001997 at 36 SPP, andPlane, from 0\.2749 to 0\.2758 between 12 and 24 SPP and from 0\.2725 to 0\.2728 between 48 and 60 SPP\. All remain in the panels below\.
### A\.3Complete 60\-SPP descriptive summary
Table[A1](https://arxiv.org/html/2610.29049#A1.T1)reports the 60\- SPP median, arithmetic mean, and sample standard deviation for every scene–method pair\. The main paper’s claims and curves use medians; the other two statistics are included to describe variation across D01–D10\.
Table A1:Combined 60\-SPP 64×\\times64 full\-frame linear\-RGB MSE summaries across D01–D10 \(n=10n=10per cell\)\. Each method cell lists median, arithmetic mean, and sample standard deviation from top to bottom\. Row\-minimum medians and means are bold; SD is not ranked\.
### A\.4Remaining ten median panels
Figures[A1](https://arxiv.org/html/2610.29049#A1.F1)–[A5](https://arxiv.org/html/2610.29049#A1.F5)report the ten scenes not shown in Figure[1](https://arxiv.org/html/2610.29049#S6.F1)\. Each compact strip places a reference overview, two visual\-only D01 native288×288288\\times 288crops per method from a1152×11521152\\times 115260\-SPP film, and the quantitative curve side by side\. The crop is not the median image\. Curve centers and numeric labels summarize D01–D1064×6464\\times 64full\-frame MSE: each curve shows the median only\. Each method’s 60\-SPP numeric annotation additionally lists raw sample SD \(denominatorn−1n\-1\)\.


Figure A1:Median MSE summaries forSphere\(top\) andLamp\(bottom\)\.

Figure A2:Median MSE summaries forPlane\(top\) andSlabs\(bottom\)\.

Figure A3:Median MSE summaries forRing\-and\-Slab\(top\) andPillars\(bottom\)\.

Figure A4:Median MSE summaries forSlab\-×1\\times 1\(top\) andSlab\-×3\\times 3\(bottom\)\.

Figure A5:Median MSE summaries forPillar\-×3\\times 3\(top\) andCornell Glass\(bottom\)\.
## Appendix BImplementation Details
The public method name isART\-MPG; immutable manifests retain historical method keys for provenance, but those keys do not denote a different algorithm\. All image metrics are computed on raw linear\-RGB OpenEXR images rather than display PNGs\.
### B\.1Feature construction, objective, and solver details
The 22\-D feature vector is, in order: receiver position \(3\), sampled\-emitter position \(3\), their displacementδ\\delta\(3\), distance \(1\), receiver normalnDn\_\{D\}\(3\), emitter normalnLn\_\{L\}\(3\), emitter directiondLd\_\{L\}\(3\), andδ^⊤nD\\hat\{\\delta\}^\{\\top\}n\_\{D\},−δ^⊤nL\-\\hat\{\\delta\}^\{\\top\}n\_\{L\}, andδ^⊤dL\\hat\{\\delta\}^\{\\top\}d\_\{L\}\. Raw positions and distance are not normalized\. The renderer setsδ^=0\\hat\{\\delta\}=0at its10−610^\{\-6\}zero\-distance guard; the loader usesδ/max\(∥δ∥,10−6\)\\delta/\\max\(\\lVert\\delta\\rVert,10^\{\-6\}\)\. Ten length logits reserve class zero; runtime sampling drops that class and renormalizes lengths 1–9\. Each R/T sigmoid probability is clipped only to machine epsilon, consistently in sampling and mass evaluation\.
Replay consists of successful solved chains exported by the online MPG\-guided manifold integrator and then aggregated and deduplicated\. For each record,ωb\\omega\_\{b\}is the normalized direction from the receiver to the first vertex of the converged chain, rather than the proposal direction before projection; this follows MPG’s sample\-centred directional construction\. Ten scenes reuse hash\-verified legacy exports, while four added scenes use fresh five\-seed exports; acquisition budgets are therefore scene\-specific rather than one uniform fourteen\-scene schedule\. IfGz\(a\)G\_\{z\}\(a\)denotes the RGB path\-contribution field before the receiver bidirectional scattering distribution function \(BSDF\),r^−1\\widehat\{r\}^\{\-1\}the export\-time reciprocal basin estimate,poffp\_\{\\rm off\}the offset\-normal density, andqN,expq\_\{N,\\rm exp\}the export length probability, the stored scalar is operationally
eb=meanRGB\(Gz\(a\)r^−1/poff\)/qN,exp\.e\_\{b\}=\\operatorname\{mean\}\_\{\\rm RGB\}\\\!\\left\(G\_\{z\}\(a\)\\widehat\{r\}^\{\-1\}/p\_\{\\rm off\}\\right\)/q\_\{N,\\rm exp\}\.\(5\)Thusebe\_\{b\}is an importance proxy rather than the full image contribution; in particular, it omits the receiver BSDF\. The Python loader preserves this field and applies only the following log transform\. Withwb=1\+log\(1\+max\(eb,0\)\)w\_\{b\}=1\+\\log\(1\+\\max\(e\_\{b\},0\)\), the exact per\-record losses are
ℓn\(b\)\\displaystyle\\ell\_\{n\}\(b\)=−logqθ\(nb∣xb\),\\displaystyle=\-\\log q\_\{\\theta\}\(n\_\{b\}\\mid x\_\{b\}\),\(6\)ℓt\(b,U\)\\displaystyle\\ell\_\{t\}\(b,U\)=−nblogqθ\(tb,U\+1∣tb,1:U,nb,xb\),U∼Unif\{0,…,nb−1\},\\displaystyle=\-n\_\{b\}\\log q\_\{\\theta\}\(t\_\{b,U\+1\}\\mid t\_\{b,1:U\},n\_\{b\},x\_\{b\}\),\\quad U\\sim\\operatorname\{Unif\}\\\{0,\\ldots,n\_\{b\}\-1\\\},\(7\)ℓω\(b\)\\displaystyle\\ell\_\{\\omega\}\(b\)=−loggθ\(ωb∣db,xb\)\.\\displaystyle=\-\\log g\_\{\\theta\}\(\\omega\_\{b\}\\mid d\_\{b\},x\_\{b\}\)\.\(8\)Training minimizes
ℒ=∑bwb\[ℓn\(b\)\+ℓt\(b,Ub\)\+λωℓω\(b\)\]∑bwb\.\\mathcal\{L\}=\\frac\{\\sum\_\{b\}w\_\{b\}\[\\ell\_\{n\}\(b\)\+\\ell\_\{t\}\(b,U\_\{b\}\)\+\\lambda\_\{\\omega\}\\ell\_\{\\omega\}\(b\)\]\}\{\\sum\_\{b\}w\_\{b\}\}\.\(9\)Because the sampled transition is multiplied bynbn\_\{b\}, its expectation is the sum of all next\-bit NLL terms in that record\.
The manifold solve uses angle\-difference constraints, at most 20 damped Newton iterations, and residual threshold10−410^\{\-4\}\. Failed surface reprojections halve the step\. The practical same\-chain predicate is intentionally weaker than the theorem’s canonical identity: it requires the primary shape sequence, packed type code, and first solved directions with\|ω1′⊤ω1−1\|<10−4\|\\omega\_\{1\}^\{\\prime\\top\}\\omega\_\{1\}\-1\|<10^\{\-4\}\.
Algorithm 2Offline prefix\-autoregressive training forART\-MPG1:Fixed replay
ℛ\\mathcal\{R\}, epoch count
EE, and direction coefficient
λω\\lambda\_\{\\omega\}\.
2:Split record groups into training/validation; initialize model and AdamW\.
3:Set
wb←1\+log\(1\+max\(eb,0\)\)w\_\{b\}\\leftarrow 1\+\\log\(1\+\\max\(e\_\{b\},0\)\)\.
4:for
e=1,…,Ee=1,\\ldots,Edo
5:foreach minibatch
ℬ\\mathcal\{B\}do
6:Evaluate length NLL from the all\-mask state\.
7:Sample one prefix per record and evaluate the scaled next\-bit NLL\.
8:Reveal the completed descriptor and evaluate vMF\-mixture NLL\.
9:Form objective[9](https://arxiv.org/html/2610.29049#A2.E9); clip gradients; update\.
10:endfor
11:Enumerate all validation prefixes; retain minimum validation loss\.
12:endfor
13:returnthe frozen proposal
qθq\_\{\\theta\}\.
### B\.2vMF density and sampling
Forκ\>0\\kappa\>0, the vMF normalizer on𝕊2\\mathbb\{S\}^\{2\}isκ/\(4πsinhκ\)\\kappa/\(4\\pi\\sinh\\kappa\), with continuous limit1/\(4π\)1/\(4\\pi\)atκ=0\\kappa=0\([Mardia and Jupp, 2000](https://arxiv.org/html/2610.29049#bib.bib9)\)\. Training evaluates the mixture NLL with log\-sum\-exp; runtime uses a stable log normalizer and linearly sums the weighted component densities\. To sample one component, it draws
w=1\+κ−1log\(u1\+\(1−u1\)e−2κ\),φ=2πu2,w=1\+\\kappa^\{\-1\}\\log\\\!\\left\(u\_\{1\}\+\(1\-u\_\{1\}\)e^\{\-2\\kappa\}\\right\),\\qquad\\varphi=2\\pi u\_\{2\},\(10\)constructs\(1−w2cosφ,1−w2sinφ,w\)\(\\sqrt\{1\-w^\{2\}\}\\cos\\varphi,\\sqrt\{1\-w^\{2\}\}\\sin\\varphi,w\), and rotates the north pole toμ\\mu\([Wood, 1994](https://arxiv.org/html/2610.29049#bib.bib8)\)\. Runtime guards the uniform draws away from the inverse\-CDF endpoints; for very smallκ\\kappait samples the sphere uniformly\. Training forms each mean asrμ/max\(∥rμ∥,10−6\)r\_\{\\mu\}/\\max\(\\lVert r\_\{\\mu\}\\rVert,10^\{\-6\}\)\. Runtime uses the corresponding unit direction when∥rμ∥2\>10−12\\lVert r\_\{\\mu\}\\rVert^\{2\}\>10^\{\-12\}and otherwise falls back to\+z\+z\. Softmax mixture weights and clamped concentrations are shared across training NLL, runtime sampling, and density evaluation; the mean conventions therefore agree outside the stated degenerate region\.
### B\.3Recorded reproducibility environment
The execution receipts bind the worker Python environment, renderer binaries, and OpenCV 4\.11\.0 EXR decoder\. The reported D01–D10 grid was rendered on one host with 80 one\-thread workers\. Binary, plugin, protocol, model, reference, and output hashes define the reproducibility boundary\. Quantitative metrics use64×6464\\times 64linear\-RGB images\.
### B\.4Result and figure provenance
The result artifact contains the complete14×6×5×10=4,20014\\times 6\\times 5\\times 10=4\{,\}200\-row matrix for D01–D10\. No per\-scene substitution, smoothing, interpolation, monotone fitting, exclusion, or replacement is applied\. The statistics artifact computes ordinary cellwise medians, arithmetic means, and sample standard deviations with denominatorn−1n\-1\. Publication curves show the median centerline without a dispersion band; 60\-SPP annotations and Table[A1](https://arxiv.org/html/2610.29049#A1.T1)retain the sample SD descriptively\. Source manifests and completion receipts bind the run rows, statistics, and publication figures by SHA\-256\. These summaries describe dispersion across the ten fixed runs and are not uncertainty intervals\.
## Appendix CTheory
### C\.1Ideal conditional estimator
Assume that𝒜\(z\)\\mathcal\{A\}\(z\)is countable,∑a∥Fz\(a\)∥<∞\\sum\_\{a\}\\lVert F\_\{z\}\(a\)\\rVert<\\infty, and every contributing chain hasqNqTRθ\>0q\_\{N\}q\_\{T\}R\_\{\\theta\}\>0\. The complete statezzis exogenous to the sampled descriptor and direction and is frozen across a primary draw and all retries;Φ¯z\\bar\{\\Phi\}\_\{z\}is deterministic; and retries are independent with exact canonical\-chain identity\. This theorem directly covers deterministic offsets\. The executed glossy\-offset routine instead draws offset normals after a seed\-dependent preliminary solve and corrects by their density; a proof for that augmented seed space is not asserted here\.
This result is the descriptor\-conditioned specialization of the SMS/MPG geometric reciprocal identity\([Zeltner et al\., 2020](https://arxiv.org/html/2610.29049#bib.bib4);[Fan et al\., 2023](https://arxiv.org/html/2610.29049#bib.bib5)\)\.
###### Theorem 1\(Descriptor\-conditioned SMS/MPG re\-hit identity\)\.
DrawD∼πθ\(⋅∣ψ\(z\)\)D\\sim\\pi\_\{\\theta\}\(\\cdot\\mid\\psi\(z\)\)andΩ0∼gθ\(⋅∣D,ψ\(z\)\)\\Omega\_\{0\}\\sim g\_\{\\theta\}\(\\cdot\\mid D,\\psi\(z\)\), and setA=Φ¯z\(D,Ω0\)A=\\bar\{\\Phi\}\_\{z\}\(D,\\Omega\_\{0\}\)\. On success, keepD,zD,zfixed and redraw directions untilΦ¯z\(D,Ω′\)=A\\bar\{\\Phi\}\_\{z\}\(D,\\Omega^\{\\prime\}\)=A\. IfJd∈\{1,2,…\}J\_\{d\}\\in\\\{1,2,\\ldots\\\}counts trials including the successful re\-hit, then the estimator in Equation[3](https://arxiv.org/html/2610.29049#S5.E3)is unbiased forI\(z\)I\(z\)\.
###### Proof\.
Conditioned onA=aA=aandD=d\(a\)D=d\(a\), each retry succeeds with probabilityRθ\(a∣d\(a\),z\)R\_\{\\theta\}\(a\\mid d\(a\),z\), soJd∼Geom\(Rθ\)J\_\{d\}\\sim\\operatorname\{Geom\}\(R\_\{\\theta\}\)and𝔼\[Jd∣A=a\]=1/Rθ\\mathbb\{E\}\[J\_\{d\}\\mid A=a\]=1/R\_\{\\theta\}\. SinceQθ\(a∣z\)=qNqTRθQ\_\{\\theta\}\(a\\mid z\)=q\_\{N\}q\_\{T\}R\_\{\\theta\},
𝔼\[I^ART\-MPG∣z\]\\displaystyle\\mathbb\{E\}\[\\widehat\{I\}\_\{\\mathrm\{ART\\text\{\-\}MPG\}\}\\mid z\]=∑dπθ\(d∣ψ\(z\)\)∑a:d\(a\)=dRθ\(a∣d,z\)Fz\(a\)πθ\(d∣ψ\(z\)\)1Rθ\(a∣d,z\)\\displaystyle=\\sum\_\{d\}\\pi\_\{\\theta\}\(d\\mid\\psi\(z\)\)\\sum\_\{a:d\(a\)=d\}R\_\{\\theta\}\(a\\mid d,z\)\\frac\{F\_\{z\}\(a\)\}\{\\pi\_\{\\theta\}\(d\\mid\\psi\(z\)\)\}\\frac\{1\}\{R\_\{\\theta\}\(a\\mid d,z\)\}=∑a∈𝒜\(z\)Fz\(a\)=I\(z\)\.\\displaystyle=\\sum\_\{a\\in\\mathcal\{A\}\(z\)\}F\_\{z\}\(a\)=I\(z\)\.Absolute summability justifies exchanging the sums and expectation\. ∎
Proposition[1](https://arxiv.org/html/2610.29049#Thmproposition1)isolates our added length\-plus\-type stratification relative to MPG’s length\-only form; repeated\-trial unbiasedness itself is inherited rather than claimed as new\.
Algorithm 3Ideal uncapped conditional estimator1:Frozen state
zz, features
x=ψ\(z\)x=\\psi\(z\), normalized proposal
qθ\(d,ω∣x\)q\_\{\\theta\}\(d,\\omega\\mid x\), filtered map
Φ¯z\\bar\{\\Phi\}\_\{z\}, and exact identity\.
2:Draw
D∼πθ\(⋅∣x\)D\\sim\\pi\_\{\\theta\}\(\\cdot\\mid x\)and
Ω0∼gθ\(⋅∣D,x\)\\Omega\_\{0\}\\sim g\_\{\\theta\}\(\\cdot\\mid D,x\)\.
3:Set
A←Φ¯z\(D,Ω0\)A\\leftarrow\\bar\{\\Phi\}\_\{z\}\(D,\\Omega\_\{0\}\); if
A=∅A=\\varnothing,return
00\.
4:Set
Jd←0J\_\{d\}\\leftarrow 0\.
5:repeat
6:
Jd←Jd\+1J\_\{d\}\\leftarrow J\_\{d\}\+1; independently draw
Ω′∼gθ\(⋅∣D,x\)\\Omega^\{\\prime\}\\sim g\_\{\\theta\}\(\\cdot\\mid D,x\)\.
7:Set
A′←Φ¯z\(D,Ω′\)A^\{\\prime\}\\leftarrow\\bar\{\\Phi\}\_\{z\}\(D,\\Omega^\{\\prime\}\)\.
8:until
A′A^\{\\prime\}has the exact canonical identity of
AA
9:return
Fz\(A\)Jd/πθ\(D∣x\)F\_\{z\}\(A\)J\_\{d\}/\\pi\_\{\\theta\}\(D\\mid x\)\.
### C\.2Finite\-work estimator boundary
Algorithm[1](https://arxiv.org/html/2610.29049#alg1)gives the implementation\-matched sampling procedure, including the offset\-density and contribution guards\. The experiment profile fixesM=2048M=2048,Cw=512C\_\{w\}=512, andCf=32C\_\{f\}=32\.
Conditional on a successful primary solve and fixed practical target, letRimplR\_\{\\rm impl\}be the probability that one independent retry satisfies the implementation predicate\. ThenJimpl∼Geom\(Rimpl\)J\_\{\\rm impl\}\\sim\\operatorname\{Geom\}\(R\_\{\\rm impl\}\)andJ~M=Jimpl𝟏\{Jimpl≤M\}\\widetilde\{J\}\_\{M\}=J\_\{\\rm impl\}\\mathbf\{1\}\\\{J\_\{\\rm impl\}\\leq M\\\}obeys
𝔼\[J~M∣A=a\]=1−\(M\+1\)\(1−Rimpl\)M\+M\(1−Rimpl\)M\+1Rimpl\.\\mathbb\{E\}\[\\widetilde\{J\}\_\{M\}\\mid A=a\]=\\frac\{1\-\(M\+1\)\(1\-R\_\{\\rm impl\}\)^\{M\}\+M\(1\-R\_\{\\rm impl\}\)^\{M\+1\}\}\{R\_\{\\rm impl\}\}\.Retry truncation alone therefore retains the fraction1−\(1−Rimpl\)M\(1\+MRimpl\)1\-\(1\-R\_\{\\rm impl\}\)^\{M\}\(1\+MR\_\{\\rm impl\}\)relative to the practical equivalence class\. This is distinct from the exact\-identity theorem\. The clamps andqNq\_\{N\}floor have distribution\-dependent effects and are fixed parts of the evaluated profile; they are not assumptions of Theorem[1](https://arxiv.org/html/2610.29049#Thmtheorem1)\.
## Appendix DExtensions and Future Work
#### Iterative guide refinement\.
ART\-MPG learns from a fixed offline replay dataset\. An extension is to alternate sampling with the current guide and teacher\-guided proposal updates, drawing inspiration from recursive on\-policy distillation\([Yin et al\., 2026](https://arxiv.org/html/2610.29049#bib.bib24)\)\. This would require a transport\-specific teacher and an objective that accounts for the changing sampling distribution\. The proposal would remain fixed during each primary draw and its re\-hit trials\.
#### Adaptive selection of training scenes\.
Training across scenes raises the question of how to allocate data collection toward configurations where the guide performs poorly\. One possible formulation is a zero\-sum interaction between a proposal learner and a scene selector, with opposing objectives based on rendering error under a fixed computational budget\. If this interaction admits an episodic Markov\-game formulation, Bayesian learning methods such as those of[Yueh et al\. \(2026\)](https://arxiv.org/html/2610.29049#bib.bib25)could provide a starting point for studying exploration under uncertainty\.
#### Decentralized allocation of sampling effort\.
A further extension is to assign separate proposal agents to different transport strata and let them allocate effort under shared computational constraints\. If their utilities depend on one another’s allocations, this yields a strategic resource\-allocation problem\. Generalized quantal response equilibria and their decentralized learning algorithms\([Zhao et al\., 2025](https://arxiv.org/html/2610.29049#bib.bib26)\)offer a possible framework for regularized allocation decisions\. Establishing suitable utilities and relating equilibrium allocations to rendering variance and runtime remain open questionsSimilar Articles
Looking Ahead and Back: Anticipatory Generative Modeling for Road-Network Trajectories
论文提出一种面向路网轨迹生成的图约束生成范式 Anticipatory Generative Modeling,并提出模型 Foresight,通过"anticipate–select–update–re-anticipate"循环解决路口转移选择问题,超越扩散模型等基线方法。
MT Forcing: Towards Stable and Consistent Multi-Turn Image Editing with Autoregressive Diffusion
论文提出 MT Forcing,通过单轮分布匹配蒸馏与教师强制引导,提升基于自回归扩散的统一多模态模型在多轮图像编辑中的视觉质量稳定性与跨轮内容一致性。
PathFM: Fast Generative Forecasting of Multivariate Trajectories
PathFM 是一个基于流匹配的轨迹式时间序列预测模型,能够快速生成完整未来路径并支持多路径采样,同时提出 fps-bench 评测基准以衡量路径级依赖结构的建模能力。
DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models
本文提出DiffusionART,一种通过聚合奖励方向进行扩散模型强化学习微调的新方法,消除了一致性力并用信任区域掩码实现免EMA调参的稳定训练,在HunyuanVideo-13B上以约12倍更少的GPU小时数取得显著更高的训练奖励。
Manifold-Guided Attention Steering
Proposes Manifold-Guided Attention Steering (MAGS), a trajectory-aware inference-time intervention that corrects reasoning errors in LLMs by projecting attention outputs back to a learned correctness manifold when deviation exceeds a threshold, outperforming static steering methods across math, code, and molecular benchmarks.