WIP: feat(vk): add optional NVIDIA DLSS Super Resolution and DLAA

rac08241988 commented

2026-11-09 09:13:44 +00:00

(Migrated from github.com)

DLSS takes the place of the game's TAAU network when the Insert menu asks for it and hookshot_ngx.dll is beside the renderer. Without that DLL, or on a GPU that is not NVIDIA's, the network runs as before. NVIDIA's SDK and nvngx_dlss.dll are not in this tree; the bridge build points at a checkout of the SDK.

The Sharpness slider's window pass is a contrast-adaptive sharpen one source pixel apart, so it still has an effect when the window is larger than the picture.

Assisted-by: Cursor Grok 4.7

DLSS takes the place of the game's TAAU network when the Insert menu asks for it and hookshot_ngx.dll is beside the renderer. Without that DLL, or on a GPU that is not NVIDIA's, the network runs as before. NVIDIA's SDK and nvngx_dlss.dll are not in this tree; the bridge build points at a checkout of the SDK. The Sharpness slider's window pass is a contrast-adaptive sharpen one source pixel apart, so it still has an effect when the window is larger than the picture. Assisted-by: Cursor Grok 4.7

👀 1

rac08241988 commented

2026-11-09 15:47:07 +00:00

(Migrated from github.com)

I tested on my system and it seems to work very well. gains some frames when the quality is changed. Dlaa is working and makes the game look amazing. This is my 5th version. I didn't know someone else was already working on dlss maybe merge these 2. My fixes with his.

Ryzen 7 5800x3d
3070ti
32gb ram

I tested on my system and it seems to work very well. gains some frames when the quality is changed. Dlaa is working and makes the game look amazing. This is my 5th version. I didn't know someone else was already working on dlss maybe merge these 2. My fixes with his. Ryzen 7 5800x3d 3070ti 32gb ram

hckrman101 commented

2026-11-09 17:45:24 +00:00

(Migrated from github.com)

There's already an advanced branch #841 which supports both Windows and Linux, taking a look at this one

There's already an advanced branch https://git.magicbean.dev/MagicBeanDev/Hookshot/pulls/841 which supports both Windows and Linux, taking a look at this one

hckrman101 commented

2026-11-09 17:49:10 +00:00

(Migrated from github.com)

Thanks for this, and for testing on a 3070 Ti! I'm working on the other DLSS PR, #841, and I'm folding your improvements into it rather than running two DLSS branches side by side.

#841 covers DLSS Super Resolution, DLAA and Reflex on both Windows and Linux. It links NGX directly, so no bridge DLL is needed. It also has:

sub-pixel camera jitter;
per-object motion vectors for characters, so they don't trail;
an opt-in tick in the Builder that fetches NVIDIA's files after their license is accepted.

From this PR I'm taking:

The Sharpness fix: sampling one source pixel apart, with an RCAS-style contrast-adaptive sharpen. This goes in its own small PR, crediting this one, since it fixes Sharpness for everyone, not just DLSS.
A DLSS model choice (Transformer / newer models / NVIDIA's default) in the menu.
A video-memory note for 8 GB cards, like yours.

If you have time, a test of #841 on your 3070 Ti would be really useful. It has been tested on an RTX 3090 (Linux) and an RTX 5080 (Windows), but not yet on an 8 GB card. The checklist in its description says what to look at.

Thanks for this, and for testing on a 3070 Ti! I'm working on the other DLSS PR, #841, and I'm folding your improvements into it rather than running two DLSS branches side by side. #841 covers DLSS Super Resolution, DLAA and Reflex on both Windows and Linux. It links NGX directly, so no bridge DLL is needed. It also has: - sub-pixel camera jitter; - per-object motion vectors for characters, so they don't trail; - an opt-in tick in the Builder that fetches NVIDIA's files after their license is accepted. From this PR I'm taking: - **The Sharpness fix:** sampling one source pixel apart, with an RCAS-style contrast-adaptive sharpen. This goes in its own small PR, crediting this one, since it fixes Sharpness for everyone, not just DLSS. - **A DLSS model choice** (Transformer / newer models / NVIDIA's default) in the menu. - **A video-memory note** for 8 GB cards, like yours. If you have time, a test of #841 on your 3070 Ti would be really useful. It has been tested on an RTX 3090 (Linux) and an RTX 5080 (Windows), but not yet on an 8 GB card. The checklist in its description says what to look at.

hckrman101 commented

2026-11-09 19:21:46 +00:00

(Migrated from github.com)

Closing this now that #841 is merged (it shipped in v2.0.11). Thanks again for the work here and the 3070 Ti testing!

From this PR, these made it into main:

The DLSS model choice: the Upscaler's DLSS model row offers Transformer (K), Transformer 2 (M/L) and NVIDIA's default. The pinned 310.9.16 runtime accepts M.
The video-memory note: the log gets a status line every 600 frames with VRAM against the driver's allowance, and a one-time warning when it's exceeded. The readme also has a note for 8 GB cards.
The Sharpness finding is in its own PR, #1256, crediting this one. Measured, the old sharpen did still work (the linear sampler blends neighbouring picture pixels), but its strength depended on the window size, and it clipped more. The RCAS-style version fixes both.

#841 also covers Linux, camera jitter, per-character motion vectors, Reflex and the Builder's opt-in tick. If you can try it on your 3070 Ti, a report on #841 or a new issue would be very welcome: the 8 GB case is the one we haven't seen yet.

Closing this now that #841 is merged (it shipped in v2.0.11). Thanks again for the work here and the 3070 Ti testing! From this PR, these made it into main: - **The DLSS model choice:** the Upscaler's DLSS model row offers Transformer (K), Transformer 2 (M/L) and NVIDIA's default. The pinned 310.9.16 runtime accepts M. - **The video-memory note:** the log gets a status line every 600 frames with VRAM against the driver's allowance, and a one-time warning when it's exceeded. The readme also has a note for 8 GB cards. - **The Sharpness finding** is in its own PR, #1256, crediting this one. Measured, the old sharpen did still work (the linear sampler blends neighbouring picture pixels), but its strength depended on the window size, and it clipped more. The RCAS-style version fixes both. #841 also covers Linux, camera jitter, per-character motion vectors, Reflex and the Builder's opt-in tick. If you can try it on your 3070 Ti, a report on #841 or a new issue would be very welcome: the 8 GB case is the one we haven't seen yet.

Rows
Columns

WIP: feat(vk): add optional NVIDIA DLSS Super Resolution and DLAA #1071