DLSSG for SM86 (proxy) - 0.4.1 Version

中文 · English

Enables NVIDIA DLSS Frame Generation (DLSS-G) on RTX 30-series (SM86) and RTX 20-series (SM75), and accelerates DLSS Super Resolution with INT8 on RTX 30 / 20. Windows x64 / D3D12; the runtime files are version.dll, dlssg_sm86.ini and the dlssg_sm86\ directory.

Upgrading from 0.3.x? Read this first: 0.4.0 is a major upgrade. The file layout and the install steps changed; do not copy it over the old version.

1. Remove the old version: quit the game, then delete every proxy DLL of this project from the game directory (version.dll, plus any winmm.dll / dbghelp.dll / dinput8.dll / dxgi.dll / d3d12.dll you placed before) and dlssg_sm86.ini.
2. Copy the three items of the new version: version.dll, dlssg_sm86.ini and the whole dlssg_sm86\ directory. The runtime, backend and kernel packs live in dlssg_sm86\; copying only the DLL does nothing.
3. Only one proxy at any time; if the game does not load version.dll, swap in one from alternatives\ (swap, do not add).
4. Use the new dlssg_sm86.ini: keys moved (Optimized and Preset are now in [DLSSG]); an old INI is still read but logs warnings, so redo your changes in the new INI.

The complete package (with the super-resolution files) is the release asset dlssg-release-0.4.1-x64.zip; for a first install follow MANUAL.en.html step by step. The 0.3.5 files are in archive\0.3.5\.

Changes in this release

0.4.1

• XeSS-FG engine (optional):
  • Multi-frame generation (MFG): [FrameGeneration] Engine=XeSS goes to 2X / 3X / 4X (MaxGeneratedFrames 1-3). The model is Intel's model 15 as ported by MFG-Unlocker 1.6.0: 2X still runs model 18's graph (4-5 % more than before), 3X / 4X run model 15's own multi-frame path with the i-th generated frame of a real frame at t = i / (N + 1). The [XeSSFG] Model key is removed (one model); an old INI that writes it gets one warning.
  • Tensor-core CUDA backend: the port's trunk layers (3x3 INT8 convolutions) are hand-written INT8 IMMA kernels (mma.sync.m16n8k16) with cp.async double buffering; the whole graph runs through this project's own D3D12 + NVAPI CUDA runner on the game's command list and never loads Intel's libxess_fg.dll.
  • [XeSSFG] FlowResolution=CoarseAt4K: Coarse automatically when the output has at least 3840 x 2160 pixels, Middle at 1440p.
  • [XeSSFG] UseHudless is now three-state: Auto (factory, decided per frame) / 1 (strict: a frame without a usable HUD-less image is refused) / 0 (never).
  • On RTX 30 it is 8-15 % faster than DLSS-G at 1440p and below, and 27-30 % slower from 4K up (see "Performance").
• [SuperResolution] Mode is now explicit presets: Off (factory) / M-INT8 / L-INT8 / M-FP16 / L-FP16; 0.4.0's Mode=On is removed, and an invalid value is configuration_invalid with super resolution not enabled.
• [DLSSG] Preset is A / B / C only: Auto is removed (it runs as A), factory A.
• Keys removed and no longer read: Optimized / Preset in their old places, [Compatibility] AccelerateNewerGpus / ForceGeneratedFrames / SM75Family / HardwareBilinear / LaunchChains / PlainVariant, [Backends], and more; writing one logs a warning and the factory value applies. RTX 40 / 50 always run this project's kernels.
• The R580 driver floor is now hard: a CUDA 13 cubin the driver refuses is an error that refuses frame generation (driver_below_floor / kernel_image_unavailable), no longer a quiet fall back to PTX.
• Super-resolution INT8 first-frame hitch fixed: the twelve weight-image uploads moved off the first observed Evaluate (that Evaluate's CPU time used to be 5-6 ms higher than the same preset in FP16).
• An explicit switch that cannot be carried out is refused: a [Runtime] Path / SRPath naming a wrong file, a missing packaged file, KernelImage=Cubin on RTX 40 / 50 and the like are refused with the reason in the log, never replaced by another path.
• Logging: one text log, dlssg_sm86\logs\dlssg_sm86.log (0.4.0's per-process frames_<pid>.jsonl is gone); [Logging] FrameTiming=1 writes frame_timing.txt.

0.4.0

• Super-resolution INT8 (RTX 30 / 20): with [SuperResolution] Mode=On, the super-resolution modes the game runs with preset L or M (Performance and Ultra Performance in most games) use this project's INT8 kernels; no preset is changed. One super-resolution evaluate takes about L −5..−17 %, M −7..−24 % GPU time against the same preset in FP16 (RTX 3070 / 3080 Ti). Off by default. The two super-resolution files are only in the complete package attached to the release (see "Install & upgrade").
• DLSS-G preset C (FG Lite, optional): [DLSSG] Preset=C, about −30..−37 % GPU time per generated frame at 2X and about −46 % at 4X, with slightly lower image quality (mostly on the HUD and at occlusion edges).
• XeSS-FG engine (optional): [FrameGeneration] Engine=XeSS generates frames with Intel's XeSS-FG model, 2X only; on RTX 30 it is slower than DLSS-G, see "Known issues". DLSS-G stays the default.
• The bundled runtime is now 310.9.16 (bit-identical output to 310.9.1 under the default preset), the only build, up to 6X; the 310.1 build is no longer provided. The factory Optimized changes from 1 to 2 (adds lossy image kernels; side by side there is no visible difference from 1).
• File layout: the runtime, backend and kernel packs moved out of the proxy DLL into the dlssg_sm86\ directory (the proxy is under 1 MB); the INI gains [DLSSG] / [XeSSFG] sections; [Logging] FrameTiming=1 logs per-frame GPU time for performance reports.

0.3.5

• Fixes a defect present since 0.3.0: after the game re-creates the frame-generation feature (menu changes, a resolution or quality change, toggling frame generation; some games re-create it often while running), the wrong optimized inference kernel could be used, seen as corrupted generated frames, and possibly as a random crash / GPU driver reset after playing for a while (issue #561 may be related). Optimized=0 is not affected. In an offline test on an RTX 3070 with 15 re-creations, the old build produced wrong generated frames in three of four runs; the new build is bit-identical in every run. On 0.3.x, just replace version.dll.

0.3.4

• Fixes the 0.3.3 crash on RTX 30 (GPU driver reset at startup or once DLSS is on; Forza Horizon 6 shows FHC01; issues #535 / #538 / #540 / #542). Cause: with the NVIDIA App DLSS override or an NGX OTA model update active, the DLSS Super Resolution model also received the GPU-architecture rewrite this project makes for Streamline, took a path that does not belong to an RTX 30 and hung the GPU. NVIDIA's own components now always get the real architecture; the rewrite applies to Streamline and the game only. On 0.3.3, just replace version.dll.

0.3.3

• RTX 20 / 30: the GPU-architecture rewrite shown to the game is now armed at game start and reports RTX 50. Games on Streamline 2.8 (e.g. Final Fantasy VII Rebirth) used to decide "this GPU does not support DLSS-G" during startup and drop the frame-generation plugin (issues #509 / #528); the rewrite is now in place before that check, and the game's 3X / 4X / 6X options unlock as on an RTX 50.
• Optimized=1 no longer skips the repeated real-frame copy within a group (SkipRepeatedRealCopy; off at every level, set it to 1 yourself if wanted). It had never been validated in a live game, and a flicker report followed on 0.3.2 (issue #532). Level 1 remains bit-identical to the official image.

0.3.2

• Part of the inference kernels rewritten (310.9 build): the generated image now matches the official DLSS-G exactly (bit-identical in tests on an RTX 3080 Ti and an RTX 5070; an RTX 2080 Ti's output is bit-identical to the 3080 Ti's), no longer lossy, with a small speed-up (0–8% on the 3080 Ti).
• Optimization levels reworked: [FrameGeneration] Optimized is now 0–3. 0 stock kernels, no acceleration; 1 every acceleration, image bit-identical to the official one (factory default); 2 adds lossy image kernels, faster, about 50 dB PSNR or better against the official image (310.9 build only); 3 everything lossy, fastest. See docs/INSTALL.en.md.

0.3.1

• Fixed frame generation not enabling on RTX 20-series (Turing). RTX 20 now works with the factory dlssg_sm86.ini like RTX 30, including 6X (310.9 build, MaxGeneratedFrames=5, game plugin permitting).
• Factory MaxGeneratedFrames is now 3 (4X); set 5 for 6X.

0.3.0

• Reverted from the native build to the proxy build. The native approach (a self-built NGX host) had game-compatibility problems that were hard to fix; this build uses a proxy DLL around the unmodified factory runtime, leaving the game's calls to NGX unchanged, which is more compatible.
• The 310.9 runtime adds 6X (MaxGeneratedFrames ceiling raised from 3 to 5). On games that themselves support Dynamic MFG, selecting "Dynamic / Auto" frame generation reaches 6X.
• Optimized kernel set (~19–32% in the offline benchmark), a two-switch factory INI, capture/replay, and diagnostic logging — see below and docs/.

Differences from the public 0.4.0 (summary)

Area	0.4.0 (public)	0.4.1 (this build)
XeSS-FG multiplier	2X (model 18)	2X / 3X / 4X (model 15 port; 2X is still model 18's graph)
XeSS-FG model choice	[XeSSFG] Model=15 / 18	one model, the key is removed
XeSS-FG backend	first D3D12 / CUDA runner	tensor-core INT8 IMMA kernels + cp.async double buffering
Driver floor	R580, refused cubin fell back to PTX	hard R580, refused cubin is an error
Super-resolution INT8 first frame	twelve weight images built during the first observed Evaluate	moved before the game starts
Configuration surface	several keys with no benefit or conflicting	cleaned up, explicit switches refused when impossible
RTX 20 / 30 numbers and performance	-	unchanged (same kernels, same bit-identical tier)

The full ledger is docs/COMPAT.md C95-C131: each entry has the symptom, root cause, the fix, and the test that pins it.

Statement & roadmap

• Kernel-level work on frame generation (DLSS-G) has essentially reached the best this project can do at this stage; kernel optimization is paused, and later releases will carry compatibility and bug fixes only.
• Super-resolution L / M INT8 ships in this release (RTX 30 / 20).
• The XeSS-FG engine (optional): the model-15 port at 2X-4X on this project's own tensor-core CUDA backend. Performance and HDR (RGBA16F) back buffers are listed under "Performance" and "Known issues".
• Vulkan support is hard to continue in the project's current state: the wrapper layer itself still has weak game compatibility and little coverage testing, and adding Vulkan on top would only bring more compatibility problems. If you need it, use one of the community patch builds.

Requirements

• OS/game: Windows 10/11 x64, D3D12.
• GPU: RTX 30-series (SM86) or RTX 20-series (Turing / SM75). Development validation and all in-game tests on an RTX 3070; the full 0.4.0 test kit also ran on an RTX 3080 Ti; RTX 20 was tested on a 2080 Ti for 0.3.x frame generation and one smoke run of this release's super-resolution INT8.
• Driver: an NVIDIA driver R580 or newer (CUDA 13.0); tested on 591.86, 610.74, 610.88 and 617.14. On an older driver frame generation is not enabled and the log has one driver_below_floor (since 0.4.1 a refused cubin no longer falls back to PTX).
• No CUDA Toolkit and no Python.

This release has a single build (bundled runtime 310.9.16, up to 6X). The two 0.3.5 builds (310.9 / 310.1) are in archive\0.3.5\.

Extra VRAM by configuration

Frame generation's own local VRAM delta (after warm-up minus before feature creation, rounded up to 10 MiB, 310.9 build, optimized). It tracks the output resolution only and is independent of the multiplier: 2X and 6X use the same amount; the multiplier costs no extra VRAM.

Aspect	Output resolution	Extra VRAM
16:9	720p	~230 MiB
16:9	1080p	~350 MiB
16:9	1440p	~540 MiB
16:9	4K / 3840×2160	~810 MiB
21:9	2560×1080	~440 MiB
21:9	3440×1440	~700 MiB
21:9	5120×2160	~1050 MiB
32:9	3840×1080	~610 MiB
32:9	5120×1440	~990 MiB
4:3	1920×1440	~430 MiB

The same height at different widths is similar (it tracks output pixel count). This is frame generation's own increment, excluding the game itself and the synthetic inputs. With Engine=XeSS the dedicated VRAM in use is about 410–530 MiB higher than with DLSS-G (RTX 3070, 4K output, five games).

Install & upgrade

For a first install, follow the manual MANUAL.en.html (Chinese MANUAL.html) at the root of the complete package step by step. In short:

1. Exit the game completely.
2. Go to the game's rendering-EXE directory (e.g. Black Myth: Wukong is ...\b1\Binaries\Win64\).
3. Copy version.dll, dlssg_sm86.ini and the whole dlssg_sm86\ directory; if a DLL of the same name already exists, back it up first. The easiest way is to download the release asset dlssg-release-0.4.1-x64.zip and unzip it: it equals this repository's root plus the two super-resolution files in dlssg_sm86\sr\ (sr_kernels.dll exceeds GitHub's per-file limit and is not in the repository; without super resolution you can skip them).
4. Launch the game, enable DLSS Frame Generation in the graphics settings, and select 2X / 3X / 4X (for 6X see "6X").
5. If the game does not load version.dll (no install record in dlssg_sm86\logs\dlssg_sm86.log with [Logging] Level=2 and frame generation on): delete it and try the proxies in alternatives\ one at a time in the order winmm.dll → dbghelp.dll → dinput8.dll → dxgi.dll → d3d12.dll; only one at any time.
6. Upgrade: exit the game, delete the old proxy DLLs (0.3.x may have placed several; delete them all), dlssg_sm86.ini and dlssg_sm86\, then copy the three new items; redo your changes in the new INI.
7. Uninstall: delete the proxy DLL, dlssg_sm86.ini and the dlssg_sm86\ directory, and put back any original DLL you backed up.

Switches in the factory dlssg_sm86.ini a player may want to change:

• [DLSSG] Optimized: optimization level 0–3. Factory 2 = every acceleration plus lossy image kernels, no visible difference from 1 side by side; 1 every acceleration, image bit-identical to the official one; 0 stock kernels, no acceleration; 3 everything lossy.
• [DLSSG] Preset: factory Auto; C = FG Lite (faster, slightly lower quality), also on this project's accelerated kernels.
• [FrameGeneration] MaxGeneratedFrames: factory 3 = 4X; 5 = 6X; the actual count is requested by the game and clamped to the runtime's ceiling.
• [FrameGeneration] Engine: factory DLSS; XeSS = Intel's XeSS-FG model (model 15; 2X runs model 18's graph): MaxGeneratedFrames 1-3 (2X-4X). [XeSSFG] FlowResolution (factory Middle, finer is slower; CoarseAt4K opt-in) and UseHudless (factory Auto) apply to XeSS only.
• [SuperResolution] Mode: factory Off; on RTX 30 / 20, On runs the game's L / M super-resolution modes on this project's INT8 kernels; within one run only the first of those modes used is accelerated. Needs dlssg_sm86\sr\ from the complete package.

Every other diagnostic/compatibility knob takes a safe default and is omitted; the full list is in docs/INSTALL.en.md. A mistyped key only falls back to its own default and logs one configuration_warning; it never turns the whole mod off.

Antivirus & signing

This project's own DLLs are code-signed: version.dll at the root, the five alternative proxies in alternatives\, and the backend and three resource packs (frame-generation kernels, super-resolution kernels, XeSS model) in dlssg_sm86\, 10 in all; NVIDIA's runtime and super-resolution files in dlssg_sm86\ keep NVIDIA's own signatures. A self-signed certificate only verifies the signer's identity and file integrity; it gives no default Windows trust, so Windows SmartScreen may still prompt "unknown publisher" on first run — that is a reputation prompt, not an antivirus detection. The certificate is self-signed as CN=DLSSG for SM86 (self-signed), SHA-1 thumbprint 85BA66762F851E49148D706915D09026281418E6; verify the signer and thumbprint via the file's Properties → Digital Signatures tab or signtool verify /pa.

Performance

Every number below comes from this project's own timing (a D3D12 timestamp before and after each Evaluate, or the [Logging] FrameTiming log). Offline GPU milliseconds measure the cost of frame generation / super resolution itself; they are not an in-game frame-rate gain.

Frame generation: real game captures (RTX 3070, offline replay)

RTX 3070, driver 617.14, the 0.4.0 test kit's full run (run_all). Inputs recorded in the games (Cyberpunk 2077, Black Myth: Wukong) replayed through the whole route (proxy → backend → NVIDIA runtime 310.9.16 or the XeSS engine); the first 4 real frames of each run dropped, median over all frames of 4 rounds. GPU milliseconds per generated frame; in brackets the change against DLSS default (negative = faster). DLSS default = the factory INI (Optimized=2, Preset=Auto); stock kernels = Optimized=0.

2X

Capture	DLSS default	Preset C	XeSS Middle	XeSS Coarse	XeSS Fine	Stock kernels
Cyberpunk 2077 1440p	2.554	1.694 (−33.7 %)	2.974 (+16.4 %)	2.344 (−8.2 %)	3.828 (+49.9 %)	3.639 (+42.4 %)
Black Myth: Wukong 1440p	2.581	1.745 (−32.4 %)	3.266 (+26.5 %)	2.634 (+2.0 %)	4.164 (+61.3 %)	3.657 (+41.7 %)
Cyberpunk 2077 4K	3.557	2.499 (−29.7 %)	6.379 (+79.3 %)	5.059 (+42.2 %)	8.296 (+133.2 %)	4.407 (+23.9 %)
Black Myth: Wukong 4K	3.756	2.679 (−28.7 %)	6.974 (+85.7 %)	5.682 (+51.3 %)	9.025 (+140.3 %)	4.560 (+21.4 %)

4X (per generated frame = the sum of a real frame's 3 Evaluates ÷ 3): Cyberpunk 2077 4K default 2.532 ms, preset C 1.392 ms (−45.0 %); Black Myth: Wukong 4K default 2.733 ms, preset C 1.517 ms (−44.5 %).

Frame generation: real game captures (RTX 3080 Ti, offline replay)

RTX 3080 Ti, driver 610.88, a full run of the same test kit (a tester's machine); method as above. During these two timing stages the GPU held 1665 MHz (below the 1950 MHz of the synthetic timing), so the absolute values are conservative; the comparisons hold.

2X

Capture	DLSS default	Preset C	XeSS Middle	XeSS Coarse	XeSS Fine	Stock kernels
Cyberpunk 2077 1440p	1.651	1.032 (−37.5 %)	2.235 (+35.4 %)	1.786 (+8.2 %)	2.791 (+69.1 %)	2.381 (+44.2 %)
Black Myth: Wukong 1440p	1.663	1.062 (−36.1 %)	2.359 (+41.9 %)	1.911 (+14.9 %)	2.936 (+76.6 %)	2.391 (+43.8 %)
Cyberpunk 2077 4K	2.171	1.473 (−32.2 %)	4.640 (+113.7 %)	3.668 (+68.9 %)	5.908 (+172.1 %)	2.860 (+31.7 %)
Black Myth: Wukong 4K	2.275	1.577 (−30.7 %)	4.763 (+109.3 %)	3.822 (+68.0 %)	6.060 (+166.3 %)	2.935 (+29.0 %)

4X: Cyberpunk 2077 4K default 1.567 ms, preset C 0.822 ms (−47.5 %); Black Myth: Wukong 4K default 1.666 ms, preset C 0.897 ms (−46.2 %).

Super-resolution INT8 against the same preset in FP16 (RTX 3070, offline)

Same test kit. GPU time of one super-resolution Evaluate in ms: median of 64 measured frames per run, median of 3 rounds. Same runtime (310.7.129), same preset; only INT8 against FP16. The two kernel families correspond to The Witcher 3 / Resonance / Forza Horizon 6 / Spider-Man (mvlo) and Black Myth: Wukong / Hogwarts Legacy (mvhi).

Family	Output	Mode	L FP16	L INT8	L change	M FP16	M INT8	M change
mvlo	3840×2160	Performance	7.443	6.459	−13.2 %	6.065	4.905	−19.1 %
mvlo	2560×1440	Performance	3.415	2.993	−12.4 %	2.787	2.294	−17.7 %
mvlo	1920×1080	Performance	1.965	1.736	−11.7 %	1.615	1.353	−16.2 %
mvlo	1920×1080	Ultra Performance	1.258	1.179	−6.3 %	1.066	0.967	−9.3 %
mvhi	3840×2160	Performance	7.305	6.127	−16.1 %	5.914	4.623	−21.8 %
mvhi	2560×1440	Performance	3.346	2.845	−15.0 %	2.719	2.171	−20.2 %
mvhi	1920×1080	Performance	1.934	1.656	−14.4 %	1.581	1.278	−19.2 %
mvhi	1920×1080	Ultra Performance	1.235	1.106	−10.4 %	1.057	0.914	−13.5 %

Super-resolution INT8 against the same preset in FP16 (RTX 3080 Ti, offline)

Same test kit and method, GPU at 1665 MHz.

Family	Output	Mode	L FP16	L INT8	L change	M FP16	M INT8	M change
mvlo	3840×2160	Performance	5.338	4.578	−14.2 %	4.341	3.455	−20.4 %
mvlo	2560×1440	Performance	2.484	2.151	−13.4 %	2.028	1.644	−18.9 %
mvlo	1920×1080	Performance	1.465	1.333	−9.0 %	1.208	1.047	−13.4 %
mvlo	1920×1080	Ultra Performance	1.000	0.948	−5.2 %	0.849	0.791	−6.9 %
mvhi	3840×2160	Performance	5.262	4.355	−17.2 %	4.252	3.239	−23.8 %
mvhi	2560×1440	Performance	2.437	2.045	−16.1 %	1.991	1.545	−22.4 %
mvhi	1920×1080	Performance	1.441	1.272	−11.8 %	1.186	0.994	−16.1 %
mvhi	1920×1080	Ultra Performance	0.979	0.892	−8.9 %	0.848	0.746	−12.0 %

In game (RTX 3070, 4K output, 2X)

RTX 3070, driver 617.14, logs of [Logging] FrameTiming=1, median of a 60 s still window after entering the world, tier 2, XeSS at Middle. Per generated frame = frame-generation GPU time; real frame rate = frames the game renders (the displayed rate is twice that).

Game	DLSS default (ms)	XeSS (ms)	XeSS real fps	Preset C (ms)	Preset C real fps
Cyberpunk 2077	3.755	6.733 (+79 %)	−8.6..−9.7 %	2.748 (−27 %)	+3.6 %
Black Myth: Wukong	4.035	7.308 (+81 %)	−11.0 %	3.085 (−24 %)	+4.3 %
The Witcher 3	3.973	7.166 (+80 %)	−15.0 %	2.974 (−25 %)	−0.8 %
A Plague Tale: Resonance	3.943	6.968 (+77 %)	−28.1 %	2.947 (−25 %)	+9.5 %
Forza Horizon 6	3.856	6.847 (+78 %)	−8.0 %	2.803 (−27 %)	+7.4 %

Cyberpunk 2077: path tracing + ray reconstruction, DLSS Ultra Performance; Black Myth: Wukong: DLSS 50 %. How the real frame rate moves depends on the game's bottleneck: in The Witcher 3 preset C saves 1 ms per generated frame and the frame rate does not change.

Super-resolution INT8 in game (RTX 3070, The Witcher 3)

RTX 3070, driver 617.14, frame generation off, frame rate over a still window after entering the world; the baseline is the same preset in FP16 (Mode=M-FP16 / L-FP16), INT8 and FP16 alternated 2–4 runs each, means in the table, gain from the medians. The preset M rows use the 0.4.0 shipped kernel sets (measured 10-10); the L rows were measured 10-09 (this family's L set has not changed since).

Output	Mode (render)	Preset	FP16 fps	INT8 fps	Gain	Per frame
3840×2160	Performance (1920×1080)	M	47.15	50.85	+7.85 %	21.21 → 19.67 ms (−1.54)
2560×1440	Performance (1280×720)	M	85.48	91.34	+6.86 %	11.70 → 10.95 ms (−0.75)
1920×1080	Performance (960×540)	M	122.14	128.35	+5.08 %	8.19 → 7.79 ms (−0.40)
3840×2160	Ultra Performance (1280×720)	L	63.56	66.13	+4.0 %	15.73 → 15.12 ms (−0.61)
2560×1440	Ultra Performance (853×480)	L	109.42	111.80	+2.2 %	9.14 → 8.94 ms (−0.19)
1920×1080	Ultra Performance (640×360)	L	145.83	147.57	+1.2 %	6.86 → 6.78 ms (−0.08)

Every gain but the last exceeds twice the same-configuration repeat spread; 1080p Ultra Performance L is within one run's spread, but all 4 INT8 runs are above all 4 FP16 runs. Image quality: preset M flickers slightly more than FP16 on still content (frame-to-frame stability of still pixels in game 61.8 vs 62.5 dB) with PSNR against ground truth no lower than FP16; preset L is within 0.6 dB of FP16.

XeSS-FG multi-frame generation (RTX 3070, offline full-route replay)

GPU milliseconds per real frame together with its generated frames (Middle, DIRECT):

Output	XeSS-FG 4X	DLSS-G default	Difference
1440p	4.5-5.0 ms	5.3-5.4 ms	8-15 % faster
4K	10.3-10.8 ms	7.9-8.5 ms	27-30 % slower

• Preset C is the fastest at every size; 2X runs model 18's graph and costs 4-5 % more than the old 2X (in game on a 5070 +1.7 %, bit-identical output).
• Black Myth: Wukong, 4K in game: 3X 32.0 fps (DLSS default 34.0), 4X 28.5 fps (DLSS default 31.5).
• These are offline GPU times, not game frame rates; the real frame rate depends on whether the bottleneck is the CPU or the GPU.

Whole generation group GPU time (RTX 3080 Ti, synthetic scene)

RTX 3080 Ti, driver 610.88, the 0.4.0 test kit (a tester's machine). GPU milliseconds of the whole frame-generation group per real frame (the span of one group's Evaluates in one submission): median of 256 groups per run, median of 4 rounds; reduction = (tier0 − current) / tier0. tier0 = stock kernels (Optimized=0), tier1 = bit-identical acceleration (1), tier2 = factory default (2). The same kit's consistency checks match the RTX 3070 reference bit for bit: frame generation 18/18, real captures 28/28, super resolution 32/32, super resolution Mode=On 4/4.

Resolution	Multiplier	tier0 (ms)	tier1 (ms)	Reduction	tier2 (ms)	Reduction
1080p	2X	1.364	0.926	32.1 %	0.917	32.8 %
1080p	3X	1.942	1.423	26.7 %	1.400	27.9 %
1080p	4X	2.520	1.914	24.0 %	1.888	25.1 %
1080p	5X	3.089	2.411	21.9 %	2.375	23.1 %
1080p	6X	3.677	2.911	20.8 %	2.867	22.0 %
1440p	2X	2.094	1.459	30.3 %	1.443	31.1 %
1440p	3X	3.030	2.238	26.1 %	2.205	27.2 %
1440p	4X	3.963	3.008	24.1 %	2.960	25.3 %
1440p	5X	4.900	3.786	22.7 %	3.719	24.1 %
1440p	6X	5.841	4.558	22.0 %	4.472	23.4 %
4K	2X	2.472	1.901	23.1 %	1.864	24.6 %
4K	3X	3.792	3.002	20.8 %	2.924	22.9 %
4K	4X	5.112	4.106	19.7 %	3.987	22.0 %
4K	5X	6.437	5.212	19.0 %	5.054	21.5 %
4K	6X	7.760	6.319	18.6 %	6.123	21.1 %

The gain is larger at lower resolution and lower multiplier (closer to launch/latency bound).

Estimating FPS after frame generation

First, in the same scene, output resolution, DLSS upscaling mode, and quality settings, turn frame generation off and read the frame rate F_off. Convert to a base frame time with 1000 / F_off, then take the whole-group generation time T_FG (ms) from the table above for that resolution, multiplier, and kernel setting.

base frame time    T_base (ms) = 1000 / F_off
group time         T_group (ms) ≈ T_base + T_FG
real-frame/group rate G (grp/s) ≈ 1000 / T_group
output frame rate  F_out (FPS)  ≈ G × M = 1000 × M / (1000 / F_off + T_FG)

M is the multiplier (2X…6X → 2…6). A group is one real frame plus M − 1 generated frames; T_FG already covers the whole group and the shared preprocessing, so do not multiply it by M − 1. With generation on, the real-frame/group rate G is lower than F_off.

Example: with frame generation off, ~50 FPS (20 ms base frame time), using the 4K 4X T_FG from the table:

Kernels	T_FG (ms)	group time (ms)	group rate (grp/s)	estimated FPS
Stock	5.112	25.112	39.8	159.3 FPS
Factory default	3.987	23.987	41.7	166.8 FPS

For other resolutions/multipliers use the matching row and your own measured F_off for that setting. This is a rough estimate — the base render time plus the generation-group cost; GPU contention, synchronization, CPU overhead, frame caps, and the monitor refresh rate all affect the real result, and the estimate is not guaranteed to equal a counter reading or the actual displayed rate.

6X

6X (5 generated frames per real frame) is NVIDIA's DLSS 4.5 Dynamic Multi Frame Generation. The bundled runtime supports it and this project makes it run on Ampere (with Engine=DLSS only); whether you get 6X depends on the game:

• The game itself supports 6X (ships a newer Streamline frame-gen plugin and offers 6X or Dynamic MFG in its menu): set MaxGeneratedFrames=5; 6X runs stably in testing.
• The game only supports 4X (ships an older 4X plugin, as most current games do): the ceiling is set by the game's plugin and this project cannot raise it to 6X. That plugin sizes its internal present queue for 4X at init, so forcing extra frames overruns it and disables frame generation or crashes. Use 4X for these games.

Real-world feedback

Black Myth: Wukong, Cyberpunk 2077, and FH6 run 4X normally in testing; Resonance A Plague Tale Legacy runs 6X by default with "Dynamic / Auto" frame generation. 0.4.0 ran DLSS default, preset C and XeSS (2X, 4K output) in these four games and The Witcher 3 on an RTX 3070 with no crash, TDR or device removal. Image quality depends on the base frame rate: at a low base rate, generated frames can show breakup and edge artifacts, more so at higher multipliers (6X is more demanding than 4X), and some games need graphics settings lowered, to raise the base rate, even at 4X for a good result. This is inherent to frame generation when there is little frame-rate headroom, not a defect of this project.

Known issues

• XeSS-FG (Engine=XeSS) is optional; the default is still DLSS-G. On RTX 30 (RTX 3070, offline full-route replay) it is 8-15 % faster than DLSS default at 1440p and below and 27-30 % slower from 4K up (see "Performance"); the cost of one generated frame is still higher than DLSS-G and the real frame rate depends on the bottleneck. About 0.4-0.5 GB more VRAM. MaxGeneratedFrames 1-3 (2X-4X); 4 or more refuses frame generation (max_generated_frames_unsupported). An RGBA16F (HDR) back buffer is refused (xefg_feature_refused), it does not switch to DLSS-G. [Debug] Capture records nothing under XeSS.
• [DLSSG] Preset=B only takes effect when the game provides both a HUD-less image and a UI plane, which most games do not; the log then has one preset_ineffective record.
• Preset C has no comparison against the official preset C on RTX 40 / 50.
• Super resolution [SuperResolution] Mode is explicit presets: Off (factory) / M-INT8 / L-INT8 / M-FP16 / L-FP16; every quality mode uses preset L / M, so Quality / Balanced / DLAA are slower than the game's own K.
• RTX 20: 310.9.16's three new kernels and XeSS were verified only by forward JIT of the sm_75 kernels on the RTX 3070 and have not run on a real Turing card; super-resolution INT8 had one smoke run on a 2080 Ti. RTX 40: the PTX frame-generation route still has no on-hardware data; RTX 50: frame generation and XeSS-FG were tested on a 5070 / 5070 Ti.

Diagnostics & boundaries

• The log is dlssg_sm86\logs\dlssg_sm86.log in the game directory: every module and every process appends to this one text file, each line starting with the local time, pid, EXE name and a level letter, then the JSON record. Default [Logging] Level=1 records errors only; use 2 or 3 when diagnosing. FrameTiming=1 also writes frame_timing.txt.
• If frame generation does nothing, check backend_*.jsonl for an install line with route active=true; if it is absent, the driver/runtime usually did not match — the reason is logged and the factory path is used.
• This release optimizes frame generation's GPU compute cost; do not read the offline time reduction as an in-game FPS gain — the real frame-rate change depends on the game and where the bottleneck is.
• All INI keys, log fields, and capture/replay are in docs/INSTALL.en.md and docs/CAPTURE.md.

SM75 source & credits

• Coldwood1026 for the RTX 20-series / SM75 adaptation (the kernel family behind the experimental SM75 route — see THIRD_PARTY_NOTICES.txt).
• NVIDIA for the DLSS-G runtime, models, and pre/post-processing (bundled, unmodified).
• Intel for the XeSS-FG shaders and model weights (MIT license, used by the XeSS engine; Intel's libxess_fg.dll is neither shipped nor loaded).
• dashdogy's RTX40MFG-Unlock v1.4.2 for the method that natively extracts the MFG (model 15) model; the model-15 port and weights this build ships were obtained that way.

License & third-party

• The project source is GPLv3.
• The bundled nvngx_dlssg.dll (310.9.16, SHA prefix cadee1aa…) and nvngx_dlss.dll (310.9.16, SHA prefix 858ccff9…), the extracted/recompiled kernel resources and the SM75 kernel family are NVIDIA and upstream third-party material, not re-licensed under GPL; the XeSS-FG shaders and weights are distributed under Intel's MIT license — see THIRD_PARTY_NOTICES.txt.