Thread-Efficient Decoding for Neural Texture Compression

Janarbek Matai; Sho Ikeda; Lukasz Lipski; Takahiro Harada

cs.CVAug 28, 2026

Thread-Efficient Decoding for Neural Texture Compression

Authors: Janarbek Matai, Sho Ikeda, Lukasz Lipski, Takahiro Harada

Organizations: Advanced Micro Devices, Inc., USA · Advanced Micro Devices, Inc., Japan · Advanced Micro Devices, Inc., Poland

Abstract

Neural texture compression (NTC) achieves higher compression ratios than BCn formats but suffers from GPU thread divergence, which significantly reduces runtime performance. In this work, we propose a shared decoder MLP architecture -- trained with a gradual decoder freezing schedule -- combined with texture clustering to reduce thread divergence by 25%-52% while preserving rendering quality. We evaluate our method on over 500 textures and multiple real rendering scenes, demonstrating up to 8.48x speedup on the Radeon RX 9070 XT GPU compared to non-shared baselines. Our key contributions include: (1) a unified shared decoder architecture that reduces divergence by grouping textures; (2) a training recipe with gradual decoder freezing that improves stability and reconstruction accuracy; (3) a semantic clustering strategy using CLIP embeddings that groups similar textures for effective decoder sharing; and (4) comprehensive performance and ablation studies validating our approach.

Figures & tables

Panel of Figure 1 — Figure 1. Top row: Rendered original scenes for Yokohama, Bistro, and ToyShop using the default textures, middle row: Rendered scenes for Yokohama, Bistro, and ToyShop using the textures generated by Clustered network, bottom row: Heatmap of divergence maps for ToyShop scene. The right side of each graph shows the thread divergence in Baseline network, and left side shows how much improved we are getting to reduce the thread divergence using Clustered–green color indicates less divergence. ToyShop have the highest 6 on divergence values as indicated by colorbar.

Figure 2. (a) Existing NTC architecture. (b) Shared decoder. (c) Similar textures share a decoder based on clustering where $s_{i}$ is size of cluster $i$ as well as number of encoders, and there are $n$ clusters

Category	Aerial	Brick	Concrete	Fabric	Floor	Metal	Plaster/Concrete	Rock	Roofing	Terrain	Wood
# of Tex.	20	80	40	37	123	24	19	50	21	107	65

Table 1. Number of textures per category.

	Baseline	Single	Clustered	Random	Category
PSNR	34.07	32.61	33.18	33.02	33.01

Table 2. Average PSNR of textures of different architectures.

	Baseline	Clustered	Single	Random
Yokohama	41.51	39.86	39.48	39.81
Bistro	48.68	47.02	45.62	46.04
ToyShop	39.40	37.47	37.37	37.24

Table 3. PSNR of three rendered images for network architectures

	Baseline	Clustered	Divergence
Yokohama	33.348	32.733	14802 / 11222 (25%)
Bistro	32.712	32.494	17941 / 13373 (25%)
ToyShop	34.220	33.687	13214 / 6407 (52%)

Table 4. PSNR of rendered scenes using Baseline vs. Clustered. The PSNR is calculated against the original rendered scene. The last column shows how much thread divergence saving we can get

Panel of Figure 3 — Figure 3. Heatmap of divergence maps for Yokohama (top) and Bistro (bottom). For ToyShop see Figure 1 . The right side of each graph shows the thread divergence in Baseline network, and left side shows how much improved we are getting to reduce the thread divergence using Clustered–green color indicates less divergence. Yokohama, Bistro and ToyShop have the highest (10 and 8) on divergence values as indicated by colorbar.

Panel of Figure 4 — Figure 4. Divergence map histogram –Baseline vs. Clustered. Top row: Baseline divergence map histogram, and the bottom row Clustered divergence map histogram. The x-axis is the block ids. The histogram shifts to left compared to top vs. bottom meaning there is less thread divergence in bottom (Clustered). The y-axis is the number of thread blocks that uses the specific texture on x-axis.

Figure 5. Is freezing decoder helps? Blue is the baseline, Organge is the network without freezing decoder, Green is the decoder freezed at epoch 175.

	GPU Freq: 1.6 GHz			GPU Freq: 2.4 GHz
Output	Shared-	Non-Shared	Non-Shared	Shared-	Non-Shared	Non-Shared
Image	Decoder	WMMA	FMA	MLP	WMMA	FMA
512²	55	259	394	37	175	270
1024²	205	1006	1530	139	678	1031
2048²	808	4002	6085	543	2691	4083
4096²	3204	16016	24315	2150	10741	16294

Table 5. Performance comparison across different configurations with 5 decoders

	GPU Freq: 1.6 GHz			GPU Freq: 2.4 GHz
Output	Shared-	Non-Shared	Non-Shared	Shared-	Non-Shared	Non-Shared
Image	Decoder	WMMA	FMA	MLP	WMMA	FMA
512²	56	503	465	38	340	313
1024²	207	1950	1833	140	1321	1244
2048²	805	7744	7357	547	5235	4976
4096²	3194	31048	29493	2161	20915	19930

Table 6. Performance comparison across different configurations with 10 decoders

Figure 6. When to freeze? a) Freezing at 25% of epoch, b) Freezing at 50% of epoch, c) Freezing at 75% of epoch. Blue is the baseline, and orange is the Shard network (100 encoders) and textures are randomly selected.

Figure 7. Ablation study of impact of MLP size for baseline vs. shared decoder. The x-axis is the MLP size with format $AxB$ where $A$ is the number of neurons, and $B$ is the number of layers.

GPU Acceleration

GPU acceleration uses parallel processors, specialized kernels, and memory-aware systems to speed computation across AI and scientific workloads.

See all 171 papers

Neural Texture Compression using Hypernetworks
Jun 25, 2026Belcour LaurentLearned Image Compression Hypernetworks
Intel Corporation
Beyond UV Mapping: Mesh Texture Compression via Surface-Aligned Texture Fields
Sep 20, 2026Jianqiang Wang, Junhui Hou, Siyu Ren +2
Department of Computer Science, City University of Hong Kong, Hong Kong SAR, China · Department of Electrical Engineering, Shanghai Jiao Tong University, Shanghai, China · Department of Computer Science and Engineering, Texas A&M University, TX 77840, US
Compression of 3D Gaussian Splatting Data Using GPU-friendly Graphics Texture Coding
Jul 16, 2026Amir Said, Randall Rauwendaal3DGS Compression 3D Gaussian Splatting
Qualcomm AI Research San Diego, CA, USA · Qualcomm Graphics Research Boulder, CO, USA

Jun 25, 2026J/K move · Enter open · S save

Neural Texture Compression using Hypernetworks

Belcour Laurent

Intel Corporation

Learned Image Compression Hypernetworks Neural Network Compression

Open paper AskPDF arXiv

Recent work on neural texture compression has demonstrated that it is possible to learn small, per-material texture representations (composed of latent textures and a small Multi-Layer Perceptron decoder) that can be decoded in real-time during shading to reproduce the input to a physically based shading model. However, existing methods require performing gradient-descent optimization per material for a given MLP and latent configuration. In this work, we train a single hypernetwork that outputs both the latent features and the MLP's weights and biases. Though the solution space is high-dimensional, this approach produces results comparable in quality to the current reference neural texture compressors. We further extend this approach to infer multiple decoders at once or even produce decoders that learn super-resolution.

Abstract

GPU Acceleration

Explore similar work

Neural Texture Compression using Hypernetworks

Beyond UV Mapping: Mesh Texture Compression via Surface-Aligned Texture Fields

Compression of 3D Gaussian Splatting Data Using GPU-friendly Graphics Texture Coding