DLSS Frame Generation Lands on RTX 30 and 20 via Proxy With 6X Support [2026]

By Indie Kings | September 20, 2026

Updated September 20, 2026: DLSSG for SM86 by sdli1995 enables NVIDIA DLSS frame generation on RTX 30 and RTX 20 cards under D3D12 through a proxy version.dll, per the official repository README. Version 0.3.5 fixes a rebuild glitch that caused garbled frames, while 310.9 builds add up to 6X where the game supports it, with optimized kernels 20 to 32 percent faster than stock on a 3080 Ti.

DLSSG for SM86 repository header card for RTX 30 and RTX 20 frame generation proxy

Image: The dlssg_for_sm86 repository header. Credit: sdli1995 via GitHub. No game footage is distributed with the project.

What the proxy enables on old cards

Per the official repository README at github.com/sdli1995/dlssg_for_sm86, the project runs DLSS-G frame generation on RTX 30 series SM86 and RTX 20 series SM75 under Windows 10 and 11 x64 with D3D12. Files are version.dll plus dlssg_sm86.ini in the game render exe folder.

The repo holds 3.6k stars, 208 forks, and 23 commits at the time of writing. Requirements are an NVIDIA driver with NGX, NVAPI, and CUDA interfaces, with 591.86 and 610.74 verified, per the README. No CUDA Toolkit or Python is needed.

Two release packs ship the same usage with different embedded runtimes, per the README. The 310.9 pack matches 310.1 at 4X and below and adds 6X support where the game allows it. The 310.1 pack is the older runtime line.

What 0.3.5 and 0.3.4 fixed

Per the README changelog, 0.3.5 fixes a defect since 0.3.0 where rebuilding frame generation after menu, resolution, or settings changes could apply the wrong optimized kernel. Symptoms were garbled generated frames and possible random crashes or driver resets. In 15 rebuilds on RTX 3070 the old build faulted 3 of 4 runs while the new build matched bit-exact every time. Version 0.3.x users should overwrite version.dll directly.

Per the README, 0.3.4 fixed a 0.3.3 crash on RTX 30 at startup or DLSS enable with driver resets, seen as FHC01 in Forza Horizon 6. Cause was the DLSS super resolution model receiving the project architecture rewrite when NVIDIA App overlay or NGX online update applied, sending RTX 30 down a wrong path. NVIDIA own components now get the true architecture while only Streamline and the game see the rewrite.

Per the README, 0.3.3 moved the architecture rewrite to game startup and reports RTX 50 so Streamline 2.8 games like Final Fantasy 7 Rebirth no longer unload the frame-gen plugin, opening 3X, 4X, and 6X options. Version 0.3.2 rewrote 310.9 kernels to bit-exact parity with official output on 3080 Ti versus 5070 and 2080 Ti versus 3080 Ti, with 0 to 8 percent speedup.

Version per README	Fix	Action
0.3.5	Rebuild uses wrong optimized kernel, garbled frames	Overwrite version.dll, Optimized 0 unaffected
0.3.4	Crash from super resolution model getting architecture rewrite	Overwrite version.dll, NVIDIA components get true arch
0.3.3	Streamline 2.8 games unload frame-gen plugin	Rewrite at startup as RTX 50, opens 3X to 6X
0.3.2	Lossy kernels replaced with bit-exact set	Optimized tiers 0 to 3, tier 1 default and bit-exact

How fast optimized kernels run

Per the README offline benchmark on RTX 3080 Ti with driver 591.86 measured September 13, 2026, optimized kernels cut frame-gen GPU cost 19 to 32 percent versus stock across 720p to 4K at 2X to 6X. The table measures GPU compute per frame group including shared preprocess, not in-game FPS gain directly.

At 1080p 2X the cost falls from 1.389 ms stock to 0.949 ms optimized, down 31.7 percent. At 4K 4X it falls from 5.577 ms to 4.442 ms, down 20.4 percent. Lower resolutions and lower multipliers gain most since startup latency dominates.

The README gives an estimation formula using base frame time plus group cost. At 50 FPS base with 20 ms frames, 4K 4X optimized at 4.442 ms implies about 163.7 FPS total versus 156.4 stock, per the worked example. Real gains depend on game bottlenecks, sync, CPU, and caps.

Resolution per README 3080 Ti	2X stock to optimized ms	4X stock to optimized ms	Cut
720p	1.135 to 0.779	2.405 to 1.839	31.4 and 23.5 percent
1080p	1.389 to 0.949	2.641 to 2.021	31.7 and 23.5 percent
1440p	2.143 to 1.491	4.163 to 3.160	30.4 and 24.1 percent
4K	2.605 to 1.987	5.577 to 4.442	23.7 and 20.4 percent

How to install and set tiers

Per the README, quit the game, copy version.dll and dlssg_sm86.ini into the render exe folder such as Black Myth Wukong b1 Binaries Win64, backing up any existing version.dll. If the game ignores version.dll, use an alternative proxy name from the alternatives folder such as winmm, dxgi, or dbghelp.

Start the game and enable DLSS frame generation at 2X, 3X, or 4X, or to 6X on 310.9 where the game plugin supports it. Upgrade by overwriting version.dll. Uninstall by restoring the backup and deleting the ini.

Factory ini keeps two decisive switches, per the README. Optimized tiers 0 to 3 run from stock with no speedup, to tier 1 fully accelerated and bit-exact as default, to tiers 2 and 3 lossy but faster. MaxGeneratedFrames defaults to 3 for 4X and moves to 5 for 6X only on 310.9. Extra VRAM runs about 350 MiB at 1080p and 810 MiB at 4K, independent of multiplier.

Optimized 0: Stock kernels, no speedup, per README.
Optimized 1: Fully accelerated, bit-exact with official, default.
Optimized 2 and 3: Lossy image kernels, faster, PSNR above 50 dB at tier 2 on 310.9.

Limits and next plans

Per the README, 6X works only where the game own plugin supports it. Games capped at 4X keep that cap because the plugin lays out its queue at init, and forcing more breaks frame-gen or crashes. Black Myth Wukong, Cyberpunk 2077, and FH6 run 4X in testing, while Resonance A Plague Tale Legacy runs 6X on dynamic auto.

Low base FPS hurts look at any multiplier, per the README, with 6X hungrier than 4X for base rate. Some titles need lower settings to lift base FPS before frame-gen looks right. That is inherent to generation with thin margins, not a project defect.

Next focus is INT8 Transformer super resolution models to lift base FPS on RTX 30 and 20, per the README. Vulkan support is paused over wrapper compatibility and test coverage. Kernel work on frame-gen is declared at its practical optimum with only compat and bug fixes ahead.

FAQ

Which cards does it support?
Per the README, RTX 30 SM86 and RTX 20 SM75 on Windows x64 D3D12. The 3080 Ti ran the full benchmark set and 2080 Ti output matched bit-exact.

How do I install it?
Per the README, copy version.dll and dlssg_sm86.ini into the render exe folder, enable frame-gen in game, and use an alternatives proxy name if the game skips version.dll.

What do Optimized tiers mean?
Per the README, 0 is stock, 1 is accelerated and bit-exact by default, 2 and 3 trade accuracy for speed.

Does it support 6X?
Per the README, only on 310.9 where the game plugin supports 6X or dynamic multi frame-gen. Games capped at 4X stay at 4X.

How much VRAM does frame-gen add?
Per the README, about 350 MiB at 1080p and 810 MiB at 4K, same for 2X through 6X.

Is it safe with antivirus?
Per the README, proxy DLLs are code-signed with a self-signed CN=DLSSG for SM86 certificate that Windows SmartScreen still flags as unknown publisher. Verify fingerprint 85BA66762F851E49148D706915D09026281418E6 via file properties or signtool.

Bottom Line

DLSSG for SM86 brings working frame generation to RTX 30 and 20 with bit-exact tier-1 kernels 20 to 32 percent cheaper than stock and 6X where games allow. Install the proxy pair, keep tier 1 default, and update version.dll for the 0.3.5 rebuild fix.