FSR 3.1 FG: HUD-less mode (RX_VK_FSR_FG_HUDLESS=1) outputs repeated real frames instead of interpolated frames #1732

Closed
opened 2026-11-19 11:08:47 +00:00 by dani10gl · 3 comments

Summary

With FSR 3.1 Frame Generation enabled (2x-4x or Auto), the FPS counter increases as expected (e.g. Base 61 -> 240 FPS), and longshot_log.txt confirms fsr3fi is running (600 made by FSR 3.1, 600 with the HUD marked), but visually every generated frame is identical to the real game frame (no real motion interpolation - looks like the base framerate).

NVIDIA DLSS FG works smoothly on the exact same setup.

Environment

  • Build: nightly-464-310e7c9d (LongshotBuilder v3.0.0)
  • GPU: NVIDIA GeForce RTX 5070 Ti (Windows 11, Vulkan)
  • Settings: 3440x1440 (internal 2580x1080), hdr=1 (Expanded HDR), Upscaler: NVIDIA DLSS or AMD FSR 3.1, Frame Gen: FSR 3.1 FG (4x)

A/B Isolation Results

  • Default (RX_VK_FSR_FG_HUDLESS=1): FPS counter shows 240 FPS, logs show 600 made by FSR 3.1 + 600 with the HUD marked, but visually frames are duplicated / not interpolated.
  • RX_VK_FSR_FG_HUDLESS=0: Interpolation immediately becomes smooth at 240 FPS, confirming fsr3fi optical flow + motion vectors work properly, though the HUD smears/artifacts (as noted in vk_framegen.cpp hud_source comments for kHudMask without HUD-less).

Suspected Root Cause

In gpu/vk/shaders/fg_hud.comp + scripts/fsr3_shaders.py (FI_PATCHES #4):

  1. fg_hud.comp checks exact equality if (scene == drawn) return shown;, otherwise writes srgb(scene) into hudless.
  2. In ffx_frameinterpolation_inpainting.h (FI_PATCHES #4), fStaticFactor is forced to 1.0f whenever abs(fCurrentInterpolationSource - fPresentColor) > 0.0.
  3. If hud.before (B10G11R11) and hud.after differ across the frame (or srgb(scene) differs by >= 1 LSB from shown), P_FI_INPAINTING overwrites the entire interpolated frame with fPresentColor (shown), turning every generated frame into a copy of the real frame.

Suggested Fix

  • Gate FI_PATCHES #4 / hudless static overlay with a small tolerance or only apply fPresentColor where the HUD mask (hud.source alpha) actually detected post-upscale UI, or
  • Investigate why scene != drawn / srgb(scene) != shown triggers across non-HUD pixels in fg_hud.comp (e.g. with hdr=1 or post-upscale passes).
### Summary With **FSR 3.1 Frame Generation** enabled (`2x`-`4x` or `Auto`), the FPS counter increases as expected (e.g. `Base 61 -> 240 FPS`), and `longshot_log.txt` confirms `fsr3fi` is running (`600 made by FSR 3.1`, `600 with the HUD marked`), **but visually every generated frame is identical to the real game frame** (no real motion interpolation - looks like the base framerate). **NVIDIA DLSS FG** works smoothly on the exact same setup. ### Environment - **Build:** `nightly-464-310e7c9d` (`LongshotBuilder v3.0.0`) - **GPU:** NVIDIA GeForce RTX 5070 Ti (Windows 11, Vulkan) - **Settings:** `3440x1440` (internal `2580x1080`), `hdr=1` (Expanded HDR), Upscaler: `NVIDIA DLSS` or `AMD FSR 3.1`, Frame Gen: `FSR 3.1 FG` (`4x`) ### A/B Isolation Results - **Default (`RX_VK_FSR_FG_HUDLESS=1`):** FPS counter shows `240 FPS`, logs show `600 made by FSR 3.1` + `600 with the HUD marked`, but visually frames are duplicated / not interpolated. - **`RX_VK_FSR_FG_HUDLESS=0`:** Interpolation immediately becomes **smooth at 240 FPS**, confirming `fsr3fi` optical flow + motion vectors work properly, though the HUD smears/artifacts (as noted in `vk_framegen.cpp` `hud_source` comments for `kHudMask` without HUD-less). ### Suspected Root Cause In `gpu/vk/shaders/fg_hud.comp` + `scripts/fsr3_shaders.py` (`FI_PATCHES` #4): 1. `fg_hud.comp` checks exact equality `if (scene == drawn) return shown;`, otherwise writes `srgb(scene)` into `hudless`. 2. In `ffx_frameinterpolation_inpainting.h` (`FI_PATCHES` #4), `fStaticFactor` is forced to `1.0f` whenever `abs(fCurrentInterpolationSource - fPresentColor) > 0.0`. 3. If `hud.before` (`B10G11R11`) and `hud.after` differ across the frame (or `srgb(scene)` differs by >= 1 LSB from `shown`), `P_FI_INPAINTING` overwrites the entire interpolated frame with `fPresentColor` (`shown`), turning every generated frame into a copy of the real frame. ### Suggested Fix - Gate `FI_PATCHES` #4 / `hudless` static overlay with a small tolerance or only apply `fPresentColor` where the HUD mask (`hud.source` alpha) actually detected post-upscale UI, or - Investigate why `scene != drawn` / `srgb(scene) != shown` triggers across non-HUD pixels in `fg_hud.comp` (e.g. with `hdr=1` or post-upscale passes).
Member

This looks like the same problem as #1310 (Game: FSR3.1 Framegen UI Glitching in Camera Motion), so I'm closing this one to keep the discussion in one place.

The reported symptom is precisely the repeated real frames from FSR 3.1 HUD-less interpolation that followed the HUD handling fix tracked by this issue and was fixed by PR #1538.

If it's different, say how and a maintainer will reopen it.

This looks like the same problem as #1310 (Game: FSR3.1 Framegen UI Glitching in Camera Motion), so I'm closing this one to keep the discussion in one place. The reported symptom is precisely the repeated real frames from FSR 3.1 HUD-less interpolation that followed the HUD handling fix tracked by this issue and was fixed by PR #1538. If it's different, say how and a maintainer will reopen it.
BeanBot 2026-11-19 11:09:16 +00:00
Author

This is not fixed by PR #1538 — my build (nightly-464-310e7c9d, built 2026-11-18, and LongshotBuilder v3.0.0) already includes PR #1538 (fg_hud.comp has hudless_colour returning shown when scene == drawn), yet the bug still reproduces 100% of the time on Windows 11 + NVIDIA RTX 5070 Ti.

Note that PR #1538 itself listed Tested on: Linux, RTX 3090 · Needs: AMD (Linux or Windows), Windows NVIDIA.

Specifically, even with #1538 applied:

  1. With default RX_VK_FSR_FG_HUDLESS=1, FSR 3.1 FG (4x, 3440x1440, hdr=1 Expanded HDR, NVIDIA DLSS or AMD FSR 3.1 upscaler) logs 600 made by FSR 3.1 + 600 with the HUD marked, but visually outputs repeated real frames (no interpolation).
  2. Setting RX_VK_FSR_FG_HUDLESS=0 immediately restores smooth 240 FPS interpolation (with the expected kHudMask-only HUD smearing from #1310).

So scene != drawn in fg_hud.comp (or a mismatch between hudless and shown) is still triggering FI_PATCHES #4 (fStaticFactor = 1.0f) across non-HUD pixels on Windows NVIDIA / hdr=1 even after #1538. Please reopen this issue for investigation.

This is **not** fixed by PR #1538 — my build (`nightly-464-310e7c9d`, built 2026-11-18, and `LongshotBuilder v3.0.0`) **already includes PR #1538** (`fg_hud.comp` has `hudless_colour` returning `shown` when `scene == drawn`), yet the bug **still reproduces 100% of the time** on **Windows 11 + NVIDIA RTX 5070 Ti**. Note that PR #1538 itself listed `Tested on: Linux, RTX 3090 · Needs: AMD (Linux or Windows), Windows NVIDIA`. Specifically, even with #1538 applied: 1. With default `RX_VK_FSR_FG_HUDLESS=1`, FSR 3.1 FG (`4x`, `3440x1440`, `hdr=1` Expanded HDR, `NVIDIA DLSS` or `AMD FSR 3.1` upscaler) logs `600 made by FSR 3.1` + `600 with the HUD marked`, but visually outputs **repeated real frames** (no interpolation). 2. Setting `RX_VK_FSR_FG_HUDLESS=0` immediately restores smooth 240 FPS interpolation (with the expected `kHudMask`-only HUD smearing from #1310). So `scene != drawn` in `fg_hud.comp` (or a mismatch between `hudless` and `shown`) is still triggering `FI_PATCHES` #4 (`fStaticFactor = 1.0f`) across non-HUD pixels on Windows NVIDIA / `hdr=1` even after #1538. Please reopen this issue for investigation.
Author

Update: I just tested the official v3.0.0 release (rc/v3.0.0) with default settings (no RX_VK_FSR_FG_HUDLESS override), and FSR 3.1 Frame Generation now interpolates smoothly (600 made by FSR 3.1, 600 with the HUD marked, 720 frames shown on time).

So this is indeed resolved in v3.0.0 — keeping the issue closed is fine. Thanks!

Update: I just tested the official **`v3.0.0` release** (`rc/v3.0.0`) with default settings (no `RX_VK_FSR_FG_HUDLESS` override), and **FSR 3.1 Frame Generation now interpolates smoothly** (`600 made by FSR 3.1`, `600 with the HUD marked`, `720 frames shown on time`). So this is indeed resolved in `v3.0.0` — keeping the issue closed is fine. Thanks!
Sign in to join this conversation.
No milestone
No project
No assignees
2 participants
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
MagicBeanDev/Longshot#1732
No description provided.