On November 11, 2026, Apex Compute announced Apex Neural Rendering (ANR), an AI-based image processing technology, and released a public beta of its frame generation feature, ANR FG. Supported GPUs include the GeForce GTX 16 series and RTX 20/30 series, as well as AMD Radeon and Intel Arc. That widens the possibility of using frame generation on GPUs older than the RTX 40/50 series that NVIDIA's official frame generation supports.
Apex Compute is also developing super-resolution technology, which upscales low-resolution images, and ray reconstruction technology, which removes ray-tracing noise. Both are still in training, however, and only the frame generation feature has been released this time. The company claims ANR can reduce processing time and GPU memory usage compared with NVIDIA DLSS. But whether the figures published by the maker can be reproduced in real gaming environments will vary with the GPU, the resolution, and how each game implements the feature. Below, we go over points to keep in mind when evaluating performance and stability, including a VRAM issue that was fixed the day after release.
Processing time up to about 26% shorter than NVIDIA DLSS, with lower VRAM usage
Frame generation creates new images between the frames the game engine actually renders, increasing the number of frames shown on screen. At 2X, one generated frame is added for each rendered frame; at 4X, three generated frames are added. While this increases the number of displayed frames, generating the images itself requires GPU compute power and memory, so processing load and VRAM usage matter.
Apex Compute compared the processing performance of ANR FG and NVIDIA DLSS on a GeForce RTX 5090 at 3840×2160 (4K). In the results the company published, ANR recorded shorter processing times than DLSS at both 2X and 4X, and also used less GPU memory.
| Measurement | ANR | DLSS | Difference | vs. DLSS |
|---|---|---|---|---|
| 2X generation time | 1.01 ms | 1.26 ms | 0.25 ms shorter | About 19.8% shorter |
| 4X average generation time | 0.61 ms | 0.82 ms | 0.21 ms shorter | About 25.6% shorter |
| GPU memory usage | 214 MB | 562 MB | 348 MB lower | About 61.9% lower |
Note: Measured by Apex Compute on an RTX 5090 at 3840×2160. Processing times are per generated frame; the 4X figure is the average across multiple generated frames. Percentages are calculated using the DLSS measurement as the baseline.
At 2X, ANR's processing time was 1.01 ms, about 19.8% shorter than DLSS's 1.26 ms. At 4X, the average processing time per generated frame was 0.61 ms, about 25.6% shorter than DLSS's 0.82 ms. GPU memory usage was 214 MB for ANR against 562 MB for DLSS, 348 MB or about 61.9% less. The company also states that ANR's model weights are 0.17 MB, but that is the size of the neural network's weights and is a separate figure from the 214 MB of GPU memory required at runtime.
These results, however, compare the time and memory required for the frame generation process itself. They do not mean that a game's overall frame rate will rise by about 20–26%. A lighter generation process may leave more headroom for other GPU work, but real game performance also depends on CPU and GPU load, graphics settings, and other factors. And even when frame generation increases the number of displayed frames, responsiveness to input does not improve by the same proportion. Judging the practical benefit requires testing that covers overall frame rate and input latency as well.
Claimed quality advantage over DLSS, but comparison conditions need care
Apex Compute has also published comparison results against NVIDIA DLSS for the quality of generated images, not just processing performance. In the company's 2X frame generation test, the image quality metric CGVQM-5 mean patch scored 96.59 for ANR and 93.66 for DLSS, putting ANR 2.93 points ahead. The 95% interval for the published difference is 1.23 to 4.85, and the company says ANR also recorded higher values on other quality metrics, such as ColorVideoVDP and PSNR, from 2X through 6X.
The published quality table, however, lacks sufficient explanation of the number of test scenes used or how the results were aggregated. In comparison images using the boat and grass from The Last of Us Part II, the ANR-side model is listed as "Model r2-w4g05-s0-w96k" and the NVIDIA SDK used for comparison as "310.9.16." It is unclear, though, whether the DLSS models and presets used in the performance evaluation and the quality comparison were identical across the board.
DLSS, the comparison target, is also being updated continuously. NVIDIA says the extended frame generation model in DLSS 4.5 improves user interface (UI) rendering quality, and advises specifying Preset B in the NVIDIA App when a game supports additional UI buffers. This is a setting for RTX 40/50 series, and it is not clear to what extent it is reflected in the comparison results Apex Compute published.
For that reason, it cannot be generalized that ANR delivers higher image quality than DLSS. The quality of generated frames can differ by scene, including fast-moving objects, fine patterns, translucent effects, and UI display. An accurate comparison requires using the same scene from the same game and matching the DLSS model, preset, and frame generation multiplier.
Broad support from GeForce GTX 16 to Radeon and Intel Arc
One of ANR FG's distinguishing features is the breadth of supported GPUs. Apex Compute lists multiple generations from NVIDIA, AMD, and Intel, as well as the Steam Deck.
| GPU maker / product | Supported series |
|---|---|
| NVIDIA GeForce | GTX 16, RTX 20/30/40/50 |
| AMD Radeon | RX 5000/6000/7000/9000 |
| Intel Arc | A series, B series |
| Handheld console | Steam Deck |
Supported operating systems are Windows 10/11, with Linux and SteamOS supported through Proton. Of particular note is that frame generation can be used even on GPUs that NVIDIA's official DLSS Frame Generation does not support, such as the GeForce GTX 16 and RTX 20/30 series. Being on the support list, however, does not guarantee sufficient performance and stability on every GPU and game.
To add it to a game, users rely on OptiScaler, which allows frame generation and super-resolution technologies to be swapped out. Apex Compute offers two versions: "Full" for games where OptiScaler is not yet installed, and "Addon" for games where it already is. Frame generation must also be enabled on the game side.
OptiScaler's developer, meanwhile, warns against using it in online games. A game's anti-cheat system may flag it as unauthorized modification, which could lead to measures such as account bans. Broad GPU support for ANR FG does not mean it can be installed safely in every game. Especially in online competitive games, check the game's terms and restrictions before installing.
A test kit for Apple silicon is available for Mac
Apex Compute also provides guidance on a Mac environment, but the form of delivery differs from the Windows version. The official site lists Apple M1 or later and macOS 26 or later as supported, and the release notes explain that it is offered as a Metal test kit for Apple silicon. With this kit, users can feed the same frames to ANR and MetalFX and compare the interpolation results and processing times.
The processing times the company measured on an M2 Max were 4.26 ms at 1080p and 15.66 ms at 4K for 2X frame generation. For 4X, the average processing time per generated frame was 2.52 ms at 1080p and 9.25 ms at 4K.
| Measured on M2 Max | 1080p | 4K |
|---|---|---|
| 2X generation time | 4.26 ms | 15.66 ms |
| 4X average generation time | 2.52 ms | 9.25 ms |
Note: Measured by Apex Compute. Processing times are per generated frame.
These results also show that frame generation processing time varies greatly with the GPU and resolution. The roughly 1 ms processing time measured on the RTX 5090 cannot be expected as is on older GPUs or Apple silicon. In addition, the Mac offering is a test kit for comparison and evaluation, and does not mean it can be installed in every game the same way as the Windows version.
VRAM growth bug fixed the day after release; caution advised on 8GB GPUs
On November 12, the day after ANR FG's release, Apex Compute released the beta "FG V.B15." This update fixes a problem in which VRAM usage kept growing every time the frame generation process was reinitialized. It affected Windows-native Direct3D 12 (D3D12) environments and occurred when frame generation was rebuilt, for example because of a resolution change or dynamic resolution adjustment.
According to the developer, before the fix, VRAM usage could pile up by hundreds of MB each time frame generation was reinitialized at 4K. If VRAM ran short as a result, the game's frame rate could drop to a few fps or the game could crash. Even if memory usage is low during normal processing, usage that grows with repeated resolution changes would interfere with long play sessions.
B15 changes things so that the copies of textures it held are released when frame generation ends or when the game stops using the texture. Frame generation instances are also actually released five seconds after the release operation. This addresses the problem of unneeded memory accumulating over repeated reinitializations.
Fixing the bug, however, does not remove the VRAM capacity constraint itself. B15's known limitations state that on a GPU with 8GB of VRAM at 4K resolution, if the game already occupies most of the memory, enabling frame generation may exceed VRAM capacity. In such environments, Apex Compute advises lowering the output resolution or texture quality.
The 214 MB GPU memory usage the company published is a value obtained in a specific measurement environment. In actual games, rendering textures, frame buffers, and other data also consume VRAM, so even if the frame generation process alone uses little memory, the game as a whole may run short of capacity. Especially on GPUs with 8GB of VRAM or less, or in environments using high-resolution, high-quality settings, it is important to check memory usage and stability with frame generation enabled.
nr-record, a tool for comparing ANR and DLSS on the same input, also released
Apex Compute has also released nr-record, a tool for comparing the image quality of ANR and DLSS under the same conditions. It records, losslessly, the input data passed from the game to the frame generation process, then replays the same recording through ANR and through the DLSS-G DLL included with the game. Matching the input conditions makes it easier to compare differences in the generated results directly.
nr-record offers synchronized playback, split-screen display, and pixel-level image magnification. For example, you can compare how the outlines of moving objects are generated in the same scene, and whether fine patterns or the UI show unnatural changes. Unlike comparing separately recorded gameplay, this has the advantage of limiting the influence of differences in input data.
Recording targets are D3D12 or Vulkan games that support DLSS frame generation. During recording, only the frames the game engine actually rendered are displayed; generated frames are not. Replaying recorded data with DLSS-G also requires a GeForce RTX 40/50 series GPU. nr-record itself contains no frame generation function, and works by using the generator being compared separately.
There are also differences in how much of the software is public. nr-record's code is available from a public repository, and OptiScaler is released under the GPL-3.0 license with source code attached to its releases. By contrast, the source code for ANR itself is slated to be released in the future, so what can be used now and what can be inspected down to the implementation do not currently match.
ANR FG is a technology that offers a new option for older GPUs that NVIDIA's official frame generation did not support. Apex Compute's measurements show it beating DLSS on both processing time and VRAM usage, but it has not been confirmed that performance and image quality in real games are uniformly better. The fact that a VRAM usage problem was fixed immediately after release also means stability still needs to be assessed at this stage, since it is a beta.
Going forward, attention will be on how much performance and image quality it delivers on older GeForce, Radeon, Intel Arc, and other GPUs. On GPUs with limited VRAM in particular, it is essential to check not only the load of frame generation but also the game's overall memory usage. To judge ANR FG's practicality, it will be important to match the DLSS model and settings used for comparison, and to verify image quality with identical input using tools such as nr-record, along with in-game stability.