OptiScaler (OptiScaler-NR-v0.9.6.zip, or OptiScaler-NR-v0.9.6-rtx40-mfg.zip with the experimental RTX 40 MFG unlock)
- Add INT8 NR models for RTX 30, 40 and 50 series GPUs on the D3D12 path: "INT8 Original" and "INT8 2xPass-in-one" (the look of two passes in one). They run wholly on kernels of this project and use no NR runtime. The files are in OptiScaler/nvfp4/int8.
- The INT8 models have been run on an RTX 50 series GPU only. They have not yet been tested on an RTX 30 or 40 series GPU.
- The NR model list is sorted by the precision of the weights: FP8, INT8, NVFP4. NVIDIA's own models are now listed as "FP8 Original" and "FP8 Optimized".
- Add a model comparison: Compare (side by side or wipe) has a new "Other side" setting that puts one pass of another NR model beside the NR picture, in place of the original picture. With "Label the sides" each side carries its model's name.
- Add 4K versions of the pass-in-one models beside the existing ones: "NVFP4 2xPass-in-one 4K", "NVFP4 3xPass-in-one 4K", "FP8 2xPass-in-one 4K" and "FP8 3xPass-in-one 4K". They give the looks of the 2xPass-in-one and 3xPass-in-one models with weights made for 4K output.
- Like the other NVFP4 and FP8 models the 4K models carry all their own weights and take none from the NR runtime, so they work with any NR runtime that offers the original model (preset 1), including 310.8.0 and 310.9.16.0.
- The FP8 models have not yet been tested on an RTX 40 series GPU.
- Suggested settings for the 2xPass-in-one and 3xPass-in-one models (NVFP4 and FP8, 4K or not): Cinematic, with Local structure lowered to 0.5 and Local tone lowered to 0.8.
Display Filter (display-filter-v0.7.0-preview.zip)
- Extract the zip and run
display_filter.exe. Keep theruntimeandnvfp4folders beside it. Supply your ownnvngx_dlssnr.dllthrough the runtime selector; it is not bundled. - The same INT8 models: "INT8: Model A" and "INT8 2xPass-in-one" (files in nvfp4/int8), and the Model list sorted by precision: FP8, INT8, NVFP4.
- Add a model comparison: "Other side of a wipe or side by side" on the Inspect and compare tab runs the same passes with a second model and shows the two models beside each other.
- The same 4K versions of the pass-in-one models.
- New "Pacing and latency" tab. Every setting on it starts at the previous behaviour:
- Presentation with DLSS FG off: "Overlay surface" (as before) or "Flip window", an ordinary window swapchain like a game's, so NVIDIA's driver Smooth Motion can be applied to display_filter.exe. Smooth Motion on the flip window has not been confirmed yet.
smooth_motion_profile.exe(beside display_filter.exe) turns Smooth Motion on (on) or off (off) for display_filter.exe only, from an administrator prompt; with no argument it only reads the setting.- Pacing: "Standard" (as before) or "Low latency": each captured frame is processed as soon as it arrives, and the newest frame is taken again right before the GPU work starts. Target FPS still limits the average rate.
- Steady edit (0 to 0.9): blends the model's change to the picture with the previous frame's, following the estimated motion, so the effect flickers less on still and slowly moving areas. Needs Estimate motion; not with foveation, passes with their own resolution, or an edit interval above 1.
- Keep still areas still (0 to 0.5): the motion estimate prefers "no movement" where it fits nearly as well.