HiMat: Diffusion-Based 4K SVBRDF Generation

• HiMat is a diffusion-based framework that generates native 4K SVBRDFs by jointly synthesizing multiple, pixel-aligned material maps such as albedo, normal, roughness, height, and metallicity.
• The architecture leverages a linear-attention DiT backbone with a lightweight CrossStitch module to ensure inter-map consistency while reducing memory and computation costs by up to 25%.
• It employs stationary wavelet transform supervision to preserve high-frequency texture details, outperforming patch-based or super-resolution approaches in material fidelity.

Searching arXiv for the exact HiMat paper and closely related material-generation references. HiMat is a diffusion-based framework for ultra-high-resolution spatially varying bidirectional reflectance distribution function (SVBRDF) generation, introduced as “HiMat: DiT-based Ultra-High Resolution SVBRDF Generation” (Wang et al., 9 Aug 2025). It is designed for native $4\text{K}$ material synthesis from text prompts, with emphasis on generating multiple pixel-aligned SVBRDF maps jointly while preserving structural coherence, high-frequency detail, and computational efficiency. The framework couples a linear-attention diffusion transformer (DiT) backbone with a lightweight CrossStitch module for inter-map consistency and stationary wavelet transform (SWT)-based supervision for detail preservation. The stated motivation is that adapting text-to-image diffusion models to multi-map SVBRDF output is difficult because consistency across maps, memory cost, and computation cost must all be controlled simultaneously (Wang et al., 9 Aug 2025).

1. Definition and problem setting

HiMat addresses ultra-high-resolution SVBRDF generation for 3D content creation (Wang et al., 9 Aug 2025). In the formulation described in the paper, the objective is not to generate a single RGB image, but to synthesize multiple aligned material maps such as albedo, normal, roughness, height, and metallicity. These maps must remain structurally matched at every pixel in order to be physically plausible when rendered (Wang et al., 9 Aug 2025).

The central challenge is that existing generative models are optimized for 3-channel RGB images, whereas SVBRDF generation requires coordinated multi-output synthesis. The paper identifies three difficulties: the need for consistency across SVBRDF maps, the high computational and memory cost associated with multi-output architectures or full self-attention, and the limitations of prior methods that either produce lower-resolution materials ($\le 512 \times 512$) or rely on progressively patch-based 4K generation, which sacrifices global coherence and detail (Wang et al., 9 Aug 2025).

This places HiMat within the broader class of material-generation systems that attempt to transfer advances in text-to-image diffusion to inverse-graphics or appearance-modeling tasks. A plausible implication is that the framework is aimed at settings where rendering quality depends not only on prompt fidelity, but also on the mutual compatibility of intrinsic material channels.

2. Architectural composition

HiMat is built around a Diffusion Transformer (DiT) backbone using a linear-attention DiT model similar to Sana [Xie et al., 2025] (Wang et al., 9 Aug 2025). The linear-attention mechanism is introduced to make native $4\text{K}$ synthesis feasible:

$$\text{LinearAttention}(Q,K,V) = \frac{\operatorname{ReLU}(Q) \left(\operatorname{ReLU}(K)^\top V\right)}{\operatorname{ReLU}(Q)\left(\operatorname{ReLU}(K)^\top \mathbf{1}\right)}$$

Unlike standard DiTs with quadratic attention, this formulation is described as reducing time and memory complexity, and the model outputs all SVBRDF maps jointly, enabling holistic, globally coherent textures (Wang et al., 9 Aug 2025).

HiMat operates in a compressed latent space using a pre-trained Deep Compression AutoEncoder (DC-AE) for SVBRDFs, with 32× spatial compression (Wang et al., 9 Aug 2025). The paper contrasts this with approaches that train dedicated VAEs from scratch, noting that HiMat avoids retraining a new VAE and instead fine-tunes an efficient pre-existing generative backbone (Wang et al., 9 Aug 2025).

A further architectural component is tileable generation through noise-rolling, allowing generated materials to be seamlessly tileable (Wang et al., 9 Aug 2025). The implementation described in the paper starts from a pre-trained Sana-1024px DiT (20 blocks, 1.6B params) and is fine-tuned progressively from 1024 $\rightarrow$ 2048 $\rightarrow$ 4096 px using MatSynth + Deschaintre datasets (6198 materials), with augmentations (Wang et al., 9 Aug 2025).

3. CrossStitch and cross-map consistency

The principal novelty of HiMat is the CrossStitch module, described as a lightweight, non-destructive module integrated into each DiT block (Wang et al., 9 Aug 2025). Its function is to learn structural correlations and alignment across the SVBRDF maps without introducing expensive cross-pixel attention or requiring major architectural redesign.

The module is designed around localized operations. It applies local $1 \times 1$ convolutions (depthwise separable), operating across the map (channel) dimension at each spatial location, and includes a parallel branch that computes mean features across maps via average pooling plus a $1 \times 1$ conv for context aggregation (Wang et al., 9 Aug 2025). The paper emphasizes that all weights are zero-initialized, so the module is initially non-operative and the DiT backbone operation is unchanged before finetuning starts, thereby preserving pretrained functionality (Wang et al., 9 Aug 2025).

Its operation is summarized as:

$$\begin{aligned} \mathbf{f} &\leftarrow \text{rearrange}(\mathbf{f}, \texttt{'m h w c -> (h w) c m'}) \ \mathbf{f} &\leftarrow \text{CrossStitch}(\mathbf{f}) \ \mathbf{f} &\leftarrow \text{rearrange}(\mathbf{f}, \texttt{'(h w) c m -> m h w c'}) \end{aligned}$$

This mechanism is motivated by the observation that SVBRDF maps are spatially aligned, unlike video or multi-view sequences. Accordingly, the framework uses local structural communication at each pixel across all output maps rather than full attention over all maps (Wang et al., 9 Aug 2025).

The paper further reports that CrossStitch adds negligible overhead and that its parameter count, memory, and inference time are drastically lower than multi-map attention or multi-branch networks (Wang et al., 9 Aug 2025). Concrete efficiency claims reported in the summary include up to 25% memory reduction and 22% lower FLOPs vs. linear attn, while attention-based models run out of memory at 4K (Wang et al., 9 Aug 2025). This suggests that the module is not merely an auxiliary fusion layer but the mechanism by which HiMat reconciles multi-map consistency with $4\text{K}$-scale tractability.

4. Wavelet supervision and high-frequency detail

HiMat employs wavelet-based fine-tuning for detail preservation, specifically using stationary wavelet transform (SWT)-based supervision (Wang et al., 9 Aug 2025). The stated reason is that high-resolution material synthesis requires preservation of high-frequency details critical for surface microstructure, and conventional losses underweight such detail.

The SWT loss is given as:

$\le 512 \times 512$0

where $\le 512 \times 512$1 enumerates LL (low-freq), LH/HL/HH (high-freq) and the loss is weighted so as to up-weight high-frequency terms (Wang et al., 9 Aug 2025). The paper notes that SWT preserves the spatial resolution of frequency components, unlike classic DWT, which downsamples (Wang et al., 9 Aug 2025).

The summary reports that ablation shows this (Tab. 4): SWT-based loss yields higher GLCM scores (texture richness) than DWT or naïve pixel losses (Wang et al., 9 Aug 2025). In this context, GLCM score is used as a measure of texture richness, providing an explicit quantitative handle on microstructural fidelity.

A plausible implication is that SWT supervision is not merely a perceptual enhancement device but a task-specific mechanism for preserving the physically relevant microstructure that SVBRDF rendering exposes under varying illumination and viewpoint.

5. Efficiency, scaling, and implementation characteristics

HiMat is described as a memory- and computation-efficient diffusion-based framework capable of generating native 4K-resolution SVBRDFs (Wang et al., 9 Aug 2025). The efficiency argument has several components. First, the linear-attention DiT scales linearly in image resolution (when sequence $\le 512 \times 512$2 embedding size $\le 512 \times 512$3), making $\le 512 \times 512$4 generation feasible on a single GPU (Wang et al., 9 Aug 2025). Second, the framework avoids expensive full self-attention and multi-branch networks (Wang et al., 9 Aug 2025). Third, CrossStitch provides cross-map communication in a lightweight form.

The implementation details reported in the paper are concise but specific:

Component	Reported detail	Source
Backbone	pre-trained Sana-1024px DiT (20 blocks, 1.6B params)	(Wang et al., 9 Aug 2025)
Training progression	1024 $\le 512 \times 512$5 2048 $\le 512 \times 512$6 4096 px	(Wang et al., 9 Aug 2025)
Data	MatSynth + Deschaintre datasets (6198 materials)	(Wang et al., 9 Aug 2025)

The runtime claim stated in the summary is that HiMat generates native 4K SVBRDFs in 90 seconds on an NVIDIA RTX 4090 (Wang et al., 9 Aug 2025). The paper also states that CrossStitch module alone enables $\le 512 \times 512$7 inference on consumer GPUs, whereas attention-based baselines often OOM at 4K (Wang et al., 9 Aug 2025).

These implementation details are central to the identity of HiMat. The framework is not presented only as a quality-improving modification to diffusion models; it is also an engineering solution to the specific problem of making multi-map $\le 512 \times 512$8 material generation practical under constrained compute and memory.

6. Empirical results and comparative positioning

The paper reports results for 4K SVBRDF generation using a large set of text prompts, stating that HiMat enables generation with strong structural coherence and high-frequency details (Wang et al., 9 Aug 2025). Its quantitative comparison against StableMaterials, which is reported as running only at $\le 512 \times 512$9, includes the following metrics (Wang et al., 9 Aug 2025):

Method	Metric	Reported value
HiMat	Aesthetics	4.43
StableMaterials	Aesthetics	3.82
HiMat	CLIPScore	24.34
StableMaterials	CLIPScore	24.14
HiMat	Human Preference Score	0.20
StableMaterials	Human Preference Score	0.19
HiMat	GLCM (texture)	1.34
StableMaterials	GLCM (texture)	0.54

The summary also states that HiMat outperforms Diffusion4K and Sana-4K in high-frequency details and faithful rendering when their images are rendered from the output SVBRDFs (Wang et al., 9 Aug 2025). In addition, super-resolving 1024px output using SUPIR is described as not as effective as native 4K HiMat generation, because SUPIR causes color bias, and doesn't recover microstructure (Wang et al., 9 Aug 2025).

Comparisons to prior approaches are sharply delineated in the source material. ControlMat is described as using patch-wise progressive generation for 4K, leading to less global consistency and loss of detail (Wang et al., 9 Aug 2025). StableMaterials is described as requiring a dedicated VAE trained on small datasets and producing low-res (512px) output so far (Wang et al., 9 Aug 2025). Multi-View/Video Methods (Wonder3D, ERA3D, etc.) are described as using full attention or multi-branch networks and as not efficient for few, spatially-aligned SVBRDF maps (Wang et al., 9 Aug 2025).

The paper’s synthesis of these comparisons is that HiMat is unique in efficiently scaling linear-attention DiT for multi-map outputs via CrossStitch, avoiding expensive cross-pixel attention or dedicated VAE retraining, and enabling native 4K generation on practical hardware (Wang et al., 9 Aug 2025). This is a comparative claim made in the source, rather than a general consensus statement across the literature.

7. Generalization, limitations, and nomenclature

Beyond text-to-material generation, the paper reports generalization to tasks such as intrinsic decomposition (Wang et al., 9 Aug 2025). In the reported experiment, HiMat (with CrossStitch) applied to SD3.5-medium backbone, trained on Hypersim dataset (albedo, normal, irradiance) shows that the CrossStitch variant improves PSNR and LPIPS across all decomposed maps compared to without CrossStitch, including comparison to the RGB$4\text{K}$0X method, although the latter is trained on larger datasets (Wang et al., 9 Aug 2025). This suggests that the architectural idea of localized cross-map coordination is not limited to SVBRDF channels.

The limitations reported in the related literature block are indirect rather than explicit for HiMat itself. The closest task-level caution in the supplied material is that super-resolution-based alternatives do not recover microstructure, and that prior multi-map or attention-heavy baselines are inefficient at 4K (Wang et al., 9 Aug 2025). A plausible implication is that HiMat’s design is motivated less by purely generative novelty than by the need to preserve physically meaningful cross-map structure under extreme resolution.

The term “HiMat” also appears elsewhere in other contexts in the supplied material, including multi-way chromatin interaction analysis, finite deformations of hyperelastic materials, hierarchical matrix methods, and human material estimation. These are distinct usages and should not be conflated with “HiMat: DiT-based Ultra-High Resolution SVBRDF Generation” (Wang et al., 9 Aug 2025). In the present sense, HiMat denotes a DiT-based material-generation framework for native $4\text{K}$1 SVBRDF synthesis, centered on CrossStitch, linear attention, and SWT-based supervision (Wang et al., 9 Aug 2025).