RealMat: Diffusion, Benchmarks & Matrix Theory

• RealMat is a multifaceted research label that includes a diffusion-based text-to-material generator producing photorealistic SVBRDF maps via a two-stage (pretraining and reinforcement learning) pipeline.
• RealMat-BaG offers a benchmark for semiconductor bandgap prediction, stressing model generalization under realistic shifts and highlighting domain robustness and fidelity challenges.
• RealMat also spans advanced matrix-variate statistics, exact real arithmetic, and real algebraic geometry, enabling novel insights in linear algebra and computational methodologies.

RealMat is a research label used in several technically distinct ways. In current arXiv literature it most explicitly denotes a diffusion-based text-to-material generator for physically based rendering workflows and, in the related name RealMat-BaG, a benchmark for experimentally grounded semiconductor bandgap prediction. In parallel, the same label is used informally for the real $(\beta=1)$ specialization of matrix-variate statistical distributions, for real-matrix problems in linear algebra, for exact real arithmetic and real radical computation, and for Real or real-algebraic structures on matrix-valued categories and rings (Zhou et al., 1 Sep 2025, Wang et al., 28 Apr 2026, Diaz-Garcia et al., 2014, Diaz-Garcia et al., 2010, Wilhelm, 2018, Lasserre et al., 2011, Cimpric, 2011, Spellmann et al., 2022).

Usage of RealMat	Central object	Principal source
Generative graphics	Text-to-material diffusion model for SVBRDF maps	(Zhou et al., 1 Sep 2025)
Materials ML evaluation	Experimental bandgap benchmark with OOD splits	(Wang et al., 28 Apr 2026)
Matrix-variate statistics	Real $(\beta=1)$ T, T-Riesz, and beta type II laws	(Diaz-Garcia et al., 2014, Diaz-Garcia et al., 2010)
Linear algebra	Real matrices as linear combinations of orthogonal matrices	(Pang, 1 Jun 2026)
Exact and algebraic computation	Expression-DAG exact real arithmetic; real radicals via moment matrices	(Wilhelm, 2018, Lasserre et al., 2011)
Real algebra and categories	Orderings on $M_n(R)$; Real matrix factorizations	(Cimpric, 2011, Spellmann et al., 2022)
Data systems analogue	Relational matrix algebra in a column store	(Dolmatova et al., 2020)

1. RealMat as a text-to-material generator

In its most explicit contemporary usage, RealMat is a text-to-material generator that maps a natural-language prompt to a full set of PBR/SVBRDF maps—including albedo or base color, height, roughness, and metallic maps—that are intended to be physically coherent and visually realistic under natural environment lighting (Zhou et al., 1 Sep 2025). The system is built on Stable Diffusion XL (SDXL) and encodes material outputs as a single $2\times 2$ RGB grid. In the reported layout, the top-left quadrant is albedo, the top-right is height or displacement, the bottom-left is roughness, and the bottom-right is metallic. At $1024\times 1024$ resolution, each map therefore occupies $512\times 512$ pixels.

The training pipeline is explicitly two-stage. First, a pretrained SDXL model is finetuned on synthetic material maps arranged in $2\times 2$ grids, so that the model retains SDXL’s text-alignment and realism priors while learning the joint distribution of material maps rather than ordinary photographs. The denoising objective is given as

$$\mathbb{E}_{(x_0,c),\,x_t,\,t}\left[\|f_\theta(x_t,t,c)-x_0\|^2\right],$$

and sampling uses a DDPM Markovian step

$$p_\theta(x_{t-1}\mid x_t,t,c)=\mathcal{N}\big(a_t x_t + b_t f_\theta(x_t,t,c),\,\sigma_t^2 \mathbf{I}\big).$$

The paper states that this stage used a synthetic SVBRDF corpus derived from approximately 8,615 material graphs, ~126k variations, and ~800k render+SVBRDF pairs, with AdamW, learning rate $2\cdot 10^{-6}$, batch size 120, and 400k iterations on 24×80GB A100 GPUs.

The second stage addresses the remaining realism gap by reinforcement learning. RealMat uses DDPO (Denoising Diffusion Policy Optimization), interpreting the denoising trajectory as a policy and maximizing

$(\beta=1)$0

with policy gradient

$(\beta=1)$1

The reward is computed only on the final sample: $(\beta=1)$2 where $(\beta=1)$3 is the SDXL VAE decoder, $(\beta=1)$4 is a Mitsuba 3 rendering pipeline, and $(\beta=1)$5 is drawn from 200 HDR environment maps from Poly Haven.

The realism reward model is itself learned. RealMat collects a large-scale dataset of realistic material images under natural lighting, uses CLIP image features, and applies a single linear layer to produce a scalar realism score. The training loss is

$(\beta=1)$6

with $(\beta=1)$7 and $(\beta=1)$8. The RL finetuning stage uses LoRA of rank 4, 50 denoising steps, learning rate $(\beta=1)$9, batch size 128, and 110 epochs on 32×80GB A100 GPUs.

Empirically, the paper reports that RL finetuning improves realism. In one user study with 641 responses, materials after finetuning were preferred in 46.8% of judgments, materials before finetuning in 28.9%, and 24.3% were marked similar. In a comparative study with 440 responses against MatFuse and ReflectanceFusion, RealMat was preferred in 53% of responses, compared with 40% for ReflectanceFusion, 3% for MatFuse, and 4% marked similar. The paper also notes limitations: prompt precision remains challenging, the $M_n(R)$0 layout imposes a resolution–channel trade-off, training is computationally expensive, and the realism reward inherits dataset and lighting biases.

2. RealMat-BaG and realistic materials evaluation

A second explicit use of the label is RealMat-BaG, where “RealMat” denotes a broader benchmark family for real-world evaluation in materials machine learning, and “BaG” denotes bandgap prediction (Wang et al., 28 Apr 2026). RealMat-BaG is framed as a deployment-centric benchmark for semiconductors under experimental labels, cross-fidelity transfer from DFT to experiment, domain-aware out-of-distribution splits, and interpretability.

The computational side of the benchmark uses 60,218 filtered Materials Project entries with PBE bandgaps and DFT-relaxed structures. The experimental side is an open-access dataset of 1,705 experimental bandgaps aligned to Materials Project structures through MPIDs. These 1,705 samples are reported as 1,183 from BandgapDatabase1, 185 from DS2, and 337 from Matbench-Expt. Bandgap ranges are excluded, multiple measurements are aggregated by the median, and when several MP polymorphs share the same reduced formula the entry with the lowest 0 K formation energy is selected.

The benchmark defines several evaluation regimes. A random split reserves 10% of the experimental dataset, or 171 samples, as a held-out test set, with the remaining 1,534 used for fine-tuning and smaller sub-regimes at 10%, 25%, and 50% of that fine-tuning pool. A feature-based OOD split uses CGCNN descriptors and k-means with $M_n(R)$1, designating the cluster whose centroid is farthest from the global mean as the OOD test set; this OOD subset contains 189 materials. A category-based leave-one-material-out split filters the data to 1,254 crystals distributed across categories such as chalcogenides, oxides, halides, antimonides, and silicides. Additional domain splits are defined over chemical systems, periodic groups, and crystal systems using MatFold.

The compared models include Linear Regression, SVR, Random Forest Regression, and several GNNs: CGCNN, CartNet, ALIGNN, CHGNet, and LEFTNet. Metrics are MAE, MRAE, and $M_n(R)$2. Under the random split, the best classical baseline reported is SVR with MRAE $M_n(R)$3, MAE $M_n(R)$4 eV, and $M_n(R)$5. Among pretrained GNNs, CHGNet reaches MRAE $M_n(R)$6, MAE $M_n(R)$7 eV, $M_n(R)$8, and LEFTNet reaches MRAE $M_n(R)$9, MAE $2\times 2$0 eV, $2\times 2$1.

The benchmark’s central result is not simply that some models perform well, but that generalization degrades sharply under realistic shifts. Errors increase under feature OOD and domain-based splits, and the paper highlights especially difficult categories such as silicides and antimonides. It further reports that pretraining on PBE data can improve data efficiency and some OOD settings, yet can also propagate systematic PBE biases. This makes RealMat-BaG less a leaderboard for absolute accuracy than a stress test of fidelity mismatch, domain robustness, and the interpretability of materials ML models.

3. RealMat as the real case of matrix-variate probability theory

In matrix-variate statistics, “RealMat” is used informally for the real case of theories developed uniformly over real normed division algebras, obtained by setting the algebra parameter to $2\times 2$2 (Diaz-Garcia et al., 2014, Diaz-Garcia et al., 2010). In this usage, matrices have entries in $2\times 2$3, $2\times 2$4 becomes $2\times 2$5, and the relevant symmetry groups are the familiar real orthogonal and Stiefel groups.

One strand concerns the generalised matricvariate $2\times 2$6-distribution, more specifically the matricvariate T-Riesz distribution. In the real case, a central type I law on $2\times 2$7 has density proportional to

$2\times 2$8

with parameter conditions

$2\times 2$9

The construction replaces the normal/Wishart pair of the classical matrix $1024\times 1024$0 law by Kotz-Riesz and Riesz distributions, in type I or type II variants. The paper states that the real T-Riesz can be constructed as $1024\times 1024$1, where $1024\times 1024$2 is real Kotz-Riesz distributed and $1024\times 1024$3 is real Riesz distributed, and that setting $1024\times 1024$4 recovers the classical real matricvariate $1024\times 1024$5 law. It also reports that $1024\times 1024$6 has a matricvariate beta-Riesz type II distribution, and that the family is orthogonally invariant and admits an affine location–scale form.

A second strand studies the distinction between matricvariate $1024\times 1024$7 and matrix multivariate $1024\times 1024$8 laws in the real case. The real central matricvariate $1024\times 1024$9 density is given as

$512\times 512$0

whereas the matrix multivariate $512\times 512$1 uses a trace kernel,

$512\times 512$2

The same paper derives the associated beta type II laws obtained from $512\times 512$3 or $512\times 512$4, and gives joint densities of singular values featuring the standard Vandermonde factor $512\times 512$5. A recurrent point across both papers is that “RealMat” here does not denote a separate probabilistic object; it denotes the real specialization of a general algebraic framework.

4. RealMat in linear algebra: sums of orthogonal matrices

In real matrix analysis, the label is used for structural questions about arbitrary real square matrices. Pang’s 2026 result resolves the question, recorded by Zhan, of how many real orthogonal matrices are needed in a real linear combination to represent an arbitrary $512\times 512$6 (Pang, 1 Jun 2026). The main theorem states that for every $512\times 512$7,

$512\times 512$8

where

$512\times 512$9

Equivalently, every real $2\times 2$0 matrix with $2\times 2$1 is a real linear combination of three real orthogonal matrices.

The paper sharpens this to an exact minimal classification: $2\times 2$2 This improves the earlier Li–Poon bound of four orthogonal matrices and matches the lower bound from Li–Zhang–Li, which showed that two orthogonal matrices do not suffice for all $2\times 2$3.

The proof combines SVD, Householder reflections, a rank-one update lemma, and a singular-value characterization of matrices that are combinations of two orthogonal matrices. The strategy is to reduce an arbitrary matrix to a nonnegative diagonal matrix via SVD, show that any such diagonal matrix is a linear combination of three orthogonal matrices, and then conjugate back. For $2\times 2$4, the decomposition is explicit: $2\times 2$5 which immediately yields a two-term representation after left and right multiplication by the orthogonal factors of the SVD.

This result is exact rather than asymptotic: three orthogonal matrices are not merely sufficient but necessary in dimensions at least three. A plausible implication is that the paper identifies a minimal “orthogonal basis template” for the whole space $2\times 2$6.

5. Exact real arithmetic and real radical computation

A computational use of the label appears in exact arithmetic and polynomial system solving. In the exact-arithmetic setting, RealMat is described as a system or library for exact real arithmetic or real algebraic computation, centered on the number type Real_algebraic and a multithreaded expression-DAG evaluation model (Wilhelm, 2018). The design follows the Exact Geometric Computation paradigm: cheap floating-point interval filters are used first, and if they do not certify a decision, the system refines a shared expression DAG using arbitrary precision arithmetic and algebraic separation bounds.

The paper describes a node type mtdag_node<Policies> storing approximations, error bounds, a requested precision, interval filters, separation bounds, parent lists, and an atomic dependency_count. Recomputation proceeds in topological order rather than through recursive descent, and parallelism is managed by a singleton Thread_manager using boost::lockfree::queue<std::function<void()>*>, std::atomic<unsigned int> thread_count, and std::atomic<unsigned int> task_count. The implementation uses MAX_THREADS = 4 and TASK_THRESHOLD = 5, and schedules node recomputation through atomic try_recomputation() calls once child dependencies are resolved. The reported motivation is not symbolic exactness for its own sake, but exact decisions such as sign tests and comparisons.

A second algebraic-computational meaning concerns the computation of real radicals of polynomial ideals through moment matrices, border bases, and semidefinite programming (Lasserre et al., 2011). The paper assumes the real variety is finite and combines border basis algorithms with dual methods based on positive semidefinite moment or Hankel matrices. Its central claim is that the resulting algorithm can compute a border basis of the input ideal and can also compute the quotient structure of the real radical ideal directly, without prior Gröbner-basis computation. In this framework, a positive finite-rank Hankel operator corresponds to a finite conic combination of point evaluations at real roots, and the kernel of such an operator is a real radical ideal. The algorithmic synthesis is therefore: border bases provide numerically stable quotient computations, while moment matrices and SDP restrict the computation to real roots and remove multiplicities.

Taken together, these two papers use “RealMat” to denote real-number exactness in two complementary senses: exact arithmetic over expression DAGs and exact extraction of the real solution set of polynomial systems.

6. Real algebra, Real matrix factorizations, and noncommutative reality conditions

In real algebraic geometry for matrix rings, the relevant object is $2\times 2$7 over a commutative unital ring $2\times 2$8, with positivity considered on the symmetric part $2\times 2$9 (Cimpric, 2011). The paper defines quadratic modules, preorderings, and orderings in $$\mathbb{E}_{(x_0,c),\,x_t,\,t}\left[\|f_\theta(x_t,t,c)-x_0\|^2\right],$$0, extends the Artin–Lang theorem and the Krivine–Stengle Positivstellensätze from $$\mathbb{E}_{(x_0,c),\,x_t,\,t}\left[\|f_\theta(x_t,t,c)-x_0\|^2\right],$$1 to $$\mathbb{E}_{(x_0,c),\,x_t,\,t}\left[\|f_\theta(x_t,t,c)-x_0\|^2\right],$$2, and proves that orderings on $$\mathbb{E}_{(x_0,c),\,x_t,\,t}\left[\|f_\theta(x_t,t,c)-x_0\|^2\right],$$3 are in one-to-one correspondence with orderings on $$\mathbb{E}_{(x_0,c),\,x_t,\,t}\left[\|f_\theta(x_t,t,c)-x_0\|^2\right],$$4 via the induction map

$$\mathbb{E}_{(x_0,c),\,x_t,\,t}\left[\|f_\theta(x_t,t,c)-x_0\|^2\right],$$5

At the same time, the paper emphasizes that this correspondence fails for preorderings, so the theory is not Morita equivalent to the scalar theory. That non-equivalence is one of the paper’s central structural claims.

A categorically different, but terminologically related, use is the theory of Real matrix factorizations (Spellmann et al., 2022). Starting from a potential $$\mathbb{E}_{(x_0,c),\,x_t,\,t}\left[\|f_\theta(x_t,t,c)-x_0\|^2\right],$$6, classical matrix factorizations are $$\mathbb{E}_{(x_0,c),\,x_t,\,t}\left[\|f_\theta(x_t,t,c)-x_0\|^2\right],$$7-graded free $$\mathbb{E}_{(x_0,c),\,x_t,\,t}\left[\|f_\theta(x_t,t,c)-x_0\|^2\right],$$8-modules with odd differential $$\mathbb{E}_{(x_0,c),\,x_t,\,t}\left[\|f_\theta(x_t,t,c)-x_0\|^2\right],$$9 satisfying $$p_\theta(x_{t-1}\mid x_t,t,c)=\mathcal{N}\big(a_t x_t + b_t f_\theta(x_t,t,c),\,\sigma_t^2 \mathbf{I}\big).$$0. The paper introduces both antilinear and contravariant Real versions, organized by Real 2-representation theory for a $$p_\theta(x_{t-1}\mid x_t,t,c)=\mathcal{N}\big(a_t x_t + b_t f_\theta(x_t,t,c),\,\sigma_t^2 \mathbf{I}\big).$$1-graded finite group. In the antilinear case the key category is

$$p_\theta(x_{t-1}\mid x_t,t,c)=\mathcal{N}\big(a_t x_t + b_t f_\theta(x_t,t,c),\,\sigma_t^2 \mathbf{I}\big).$$2

the homotopy fixed point category of a Real action. In the contravariant case the equivariant matrix factorization category $$p_\theta(x_{t-1}\mid x_t,t,c)=\mathcal{N}\big(a_t x_t + b_t f_\theta(x_t,t,c),\,\sigma_t^2 \mathbf{I}\big).$$3 acquires a canonical dg duality structure, aligning the theory with Grothendieck–Witt theory and with the parity symmetries of Landau–Ginzburg orientifolds.

The main theorems are Real analogues of Knörrer periodicity. In the antilinear setting, the paper proves a Real equivariant quasi-equivalence after adding a quadratic form $$p_\theta(x_{t-1}\mid x_t,t,c)=\mathcal{N}\big(a_t x_t + b_t f_\theta(x_t,t,c),\,\sigma_t^2 \mathbf{I}\big).$$4, structurally analogous to $$p_\theta(x_{t-1}\mid x_t,t,c)=\mathcal{N}\big(a_t x_t + b_t f_\theta(x_t,t,c),\,\sigma_t^2 \mathbf{I}\big).$$5-periodicity for $$p_\theta(x_{t-1}\mid x_t,t,c)=\mathcal{N}\big(a_t x_t + b_t f_\theta(x_t,t,c),\,\sigma_t^2 \mathbf{I}\big).$$6-theory. In the contravariant setting, it proves quasi-equivalences of dg categories with duality after adding $$p_\theta(x_{t-1}\mid x_t,t,c)=\mathcal{N}\big(a_t x_t + b_t f_\theta(x_t,t,c),\,\sigma_t^2 \mathbf{I}\big).$$7 or, more canonically, $$p_\theta(x_{t-1}\mid x_t,t,c)=\mathcal{N}\big(a_t x_t + b_t f_\theta(x_t,t,c),\,\sigma_t^2 \mathbf{I}\big).$$8, structurally analogous to $$p_\theta(x_{t-1}\mid x_t,t,c)=\mathcal{N}\big(a_t x_t + b_t f_\theta(x_t,t,c),\,\sigma_t^2 \mathbf{I}\big).$$9-periodicity for Grothendieck–Witt theory. The framework also incorporates finite-group equivariance and discrete torsion twists.

A common misconception would be to treat these algebraic usages as mere reformulations of ordinary matrix theory over $2\cdot 10^{-6}$0. The literature instead uses “Real” in Atiyah’s and Grothendieck–Witt’s senses: involutive, duality-bearing, and equivariant structures, not just real-number entries.

7. Systems-level analogue: relational matrix algebra

At the systems level, work “in the spirit” of a hypothetical RealMat appears in Relational Matrix Algebra (RMA), proposed as a principled logical unification of relations and linear algebra inside a column-store database (Dolmatova et al., 2020). RMA extends relational algebra with matrix operations such as multiplication, inversion, QR, and SVD, but remains closed: inputs and outputs are always relations, and no separate logical matrix data type is introduced.

The core abstraction splits a relation into an order schema and an application schema. Matrix constructors $2\cdot 10^{-6}$1 and $2\cdot 10^{-6}$2 extract ordered relational views corresponding to row identities and numerical payload, while a relation constructor $2\cdot 10^{-6}$3 turns matrix rows back into tuples. On that basis, unary operators such as

$2\cdot 10^{-6}$4

and binary operators such as matrix multiplication are defined entirely relationally. The paper proves matrix consistency for the resulting operators and formalizes row origins and column origins, which preserve contextual information across operations such as transpose or covariance.

The implementation target is MonetDB, where relations are stored as columnar BATs and matrix operations can be executed either natively over BATs or via Intel MKL after physical conversion to contiguous arrays. The reported experiments show that this design performs well on mixed workloads involving joins, aggregations, regression, covariance computation, and element-wise matrix operations.

Although RMA is not itself called RealMat, it is directly relevant to the broader research pattern captured by that label: it treats matrices over real-valued data as first-class objects without abandoning schema, provenance, or relational optimization. In that sense it is a systems-level counterpart to the mathematical and statistical RealMat usages surveyed above.