Update v0.9.2 Beta

LibNTC

• Fixed the FileUnrecognized error that happened in some cases when the binary chunk size was not a multiple of 4.

Command-Line Tool

• Added --keepFileNames option that preserves the full original texture file names in the NTC file.
• Fixed the --matchBcPsnr feature on Vulkan.
• Fixed Vulkan initialization errors when using the tool over SSH.

Known Issues

Please see the Known Issues section of the main Readme file.

Assets and Drivers

Please see the release notes for v0.9.0.

Update v0.9.1 Beta

LibNTC

• Improved accuracy of DP4a based decompression and made it more consistent across GPUs and graphics APIs.

Command-Line Tool

• Added --no-dithering option that disables color dithering when saving images after graphics API decompression.

Testing Infrastructure

• Fixed reporting of PSNR values for mip chains in ImageDiff.
• Updated the BC transcoding tests to use much lower tolerances.

Other

• Updated the Donut framework to reduce the number of external library dependencies.

Known Issues

Please see the Known Issues section of the main Readme file.

Assets and Drivers

Please see the release notes for v0.9.0.

Update v0.9.0 Beta

Breaking Changes

• Changed the shape of the image decoder network (Multi-Layer Perceptron, MLP) to a smaller one.
  • This change significantly improves inference performance at a minor quality cost (about -0.25 dB PSNR).
• Replaced the BC7 encoding optimization solution to store mode and partition values for each block.
  • This change makes the BC7 encoding process 2X-6X faster at the expense of about 0.3 bits per pixel per BC7 texture in storage (when GDeflate compression is used).
• NTC files compressed with earlier versions of LibNTC are incompatible with version 0.9.0.

LibNTC

• Added support for changing the number and sizes of the MLP hidden layers individually with just a few constants.
• Added support for compressing the BC7 mode buffers and slices of the latent texture with GDeflate.
• Added support for loading NTC files using smaller MLP hidden layers than what the library is compiled for.
• Added functions for decompressing GDeflate-compressed data on the CPU using libdeflate and on the GPU using VK_NV_memory_decompression.
  • Note: GDeflate decompression on DX12 is also supported through DirectStorage, but that has to be implemented on the application side.
• Added CMake option NTC_DEBUG_SHADERS to make profiling easier.
• Refactored the code to explicitly use the same weight layout definition structures in all places.

Explorer

• Added a mode for comparing two image files to each other when the --compare command line option is used.
• Added display of per-texture and per-channel PSNR values in the results window.
• Added support for saving the manifest file including all changes made in the GUI.
• Added support for specifying per-channel loss function scales in the GUI and from the manifest.
• Changed the default graphics API to Vulkan.

Rendering Sample

• Added decompression of GDeflate-compressed buffers on the GPU with VK_NV_memory_decompression or DirectStorage (disabled by default, use --gpuGDeflate).

Command-Line Tool

• Added decompression of GDeflate-compressed buffers on the GPU for testing (disabled by default, use --gpuGDeflate).
• Added support for reading the manifest from STDIN.
• Added support for saving the manifest file automatically generated from input image files.
• Added support for specifying per-channel loss function scales from the manifest.

Testing Infrastructure

• Added ImageDiff, a small command line tool for comparing multiple pairs of images, including block compressed DDS textures.
• Added arguments to test.py to disable execution of some tests.
• Added functional tests for BC compression.

Known Issues

Please see the Known Issues section of the main Readme file.

Intel Sponza Asset

Since version 0.7.2, we provide the model and texture files for the Intel Sponza scene compatible with the RTXNTC Rendering sample as a separate download. We recommend using this model for benchmarking and not the default FlightHelmet model. In this release, the texture files are updated with the new NTC latent structure.

Note that we only provide the NTC-specific files. They are enough to run the Renderig sample in the NTC mode, but not in the reference mode. To get the original textures, please download the Sponza Base Scene and Colorful Curtains packages from the Intel GPU Research Samples page and extract them to the same folder with the NTC files, and use the --referenceMaterials argument to ntc-renderer.exe.

Shader Model 6.9 Preview Drivers

For Cooperative Vector support on NVIDIA GPUs, please use the NVIDIA Graphics Driver preview version 590.26 or newer for DX12, or at least version 570 for Vulkan. The preview drivers can be downloaded using the following links (require an NVIDIA Developer Program account):

GeForce GPUs: https://developer.nvidia.com/downloads/shadermodel6-9-preview-driver
Quadro GPUs: https://developer.nvidia.com/downloads/assets/secure/shadermodel6-9-preview-driver-quadro

Update v0.8.0 Beta

Breaking Changes

• Replaced the two latent arrays that used custom bit packing (read through a ByteAddressBuffer) with a single BGRA4 texture array (sampled through a Texture2DArray).
  • This change significantly improves inference performance and simplifies the shader code at the expense of a minor quality loss.
  • The integrations of Inference on Load and Inference on Sample need to be updated.
• Removed the network version selection, only one configuration is available now.
• Removed the legacy inference shader versions that didn't use DP4a and the versions that used CoopVec with a full Int8 pipeline. FP8 inference is now the default, and the FP8 weights are always produced during compression.
• Removed the functions enabling operation with partially loaded latent images.
• NTC files compressed with earlier versions of LibNTC are imcompatible with version 0.8.0.

LibNTC

• Added support for specifying loss function scales per-channel.
• Added header file <libntc/shaders/Bindings.h> declaring all binding indices for the compute shaders provided by the library.
• Fixed decompression errors on small images with unusual dimensions, e.g. 94x94.
• Fixed the compiler warnings on template usage in CUDA code (#3)
• Fixed the duplicate copies of weights (#4)
• Fixed the grid dimensions in the quantizer (#2)
• Improved optimizer performance.
• Redefined the known latent shapes to fit the new latent storage configurations available with BGRA4 latents.
• Simplified the math for positional encoding.
• Switched the CoopVec shader builds on DXIL to use DXC directly.
• Updated the compilers: DXC to v1.8.2505.1, Slang to v2025.18.2.

Explorer

• Added support for calculating the compression ratio over BCn formats specified in the manifest.
• Added sliders for loss function scaling per-texture.
• Show compression progress when the window is out of focus.

Command-Line Tool

• Fixed BC compression for use cases when the data comes from image files, not NTC.

Intel Sponza Asset

Since version 0.7.2, we provide the model and texture files for the Intel Sponza scene compatible with the RTXNTC Rendering sample as a separate download. We recommend using this model for benchmarking and not the default FlightHelmet model. In this release, the texture files are updated with the new NTC latent structure.

Shader Model 6.9 Preview Drivers

For Cooperative Vector support on NVIDIA GPUs, please use the NVIDIA Graphics Driver preview version 590.26 or newer for DX12, or at least version 570 for Vulkan. The preview drivers can be downloaded using the following links (require an NVIDIA Developer Program account):

GeForce GPUs: https://developer.nvidia.com/downloads/shadermodel6-9-preview-driver
Quadro GPUs: https://developer.nvidia.com/downloads/assets/secure/shadermodel6-9-preview-driver-quadro

Update v0.7.2 Beta

LibNTC

• Fixed multiple causes for crashes that happened when attempting compression of very large texture sets.
• Improved inference performance with Cooperative Vectors on DX12.

Rendering Sample

• Added plots of frame and render time and tile counts.
• Improved performance of the NTC On Feedback mode by implementing stochastic feedback.
• Moved the DLSS integration into the Donut framework.

Command-Line Tool

• Fixed integer overflows when compressing and decompressing very large texture sets.

Intel Sponza Asset

This release includes the model and texture files for the Intel Sponza scene compatible with the RTXNTC Rendering sample as a separate download. We recommend using this model for benchmarking and not the default FlightHelmet model.

Shader Model 6.9 Preview Drivers

For Cooperative Vector support on NVIDIA GPUs, please use the NVIDIA Graphics Driver preview version 590.26 or newer for DX12, or at least version 570 for Vulkan. The preview drivers can be downloaded using the following links (require an NVIDIA Developer Program account):

GeForce GPUs: https://developer.nvidia.com/downloads/shadermodel6-9-preview-driver
Quadro GPUs: https://developer.nvidia.com/downloads/assets/secure/shadermodel6-9-preview-driver-quadro

Update v0.7.1 Beta

Shader Model 6.9 Preview Drivers

For Cooperative Vector support on NVIDIA GPUs, please use the NVIDIA Graphics Driver preview version 590.26 or newer for DX12, or at least version 570 for Vulkan. The preview drivers can be downloaded using the following links (require an NVIDIA Developer Program account):

GeForce GPUs: https://developer.nvidia.com/downloads/shadermodel6-9-preview-driver
Quadro GPUs: https://developer.nvidia.com/downloads/assets/secure/shadermodel6-9-preview-driver-quadro

LibNTC

• Fixed Cooperative Vector inference on Intel Arc B-series GPUs.
• Minor code improvements.

Update v0.7.0 Beta

Shader Model 6.9 Preview Drivers

For Cooperative Vector support on NVIDIA GPUs, please use the NVIDIA Graphics Driver preview version 590.26 or newer for DX12, or at least version 570 for Vulkan. The preview drivers can be downloaded using the following links (require an NVIDIA Developer Program account):

GeForce GPUs: https://developer.nvidia.com/downloads/shadermodel6-9-preview-driver
Quadro GPUs: https://developer.nvidia.com/downloads/assets/secure/shadermodel6-9-preview-driver-quadro

LibNTC

• Switched the DX12 Cooperative Vector implementation to use the Agility SDK Preview instead of the NVIDIA custom extensions.
• Moved the Cooperative Vector weight layout conversions to happen on the GPU.
• Added support for shuffling inference output channels to make channel mappings more flexible.
• Improved code quality around inference weight layouts.

Rendering Sample

• Implemented a custom GLTF extension NV_texture_swizzle to define the NTC storage for materials.
• Improved the Inference on Feedback mode to transcode tiles in batches.
• Improved the Inference on Sample mode by replacing conditional texture channel usage with constant output channels.
• Added a display of the inference math versions that are being used for materials.

Update v0.6.1 Beta

• Improved the Inference on Feedback mode to add support for standby tiles.
• Fixed the rendering mode display in the Renderer sample in the reference mode.
• Implemented handling of VK_SUBOPTIMAL_KHR in the SDK apps (#3)

Update v0.6.0 Beta

This release adds the Inference on Feedback mode in the Renderer sample app (DX12 only).

Initial release v0.5.0 Beta

This is the first release of the new RTX Neural Texture Compression SDK.

Please observe the pre-release feature warning in the main Readme file.