Game: FSR 3.1 frame generation shows swinging limbs as see-through copies in made frames #1236

Open
opened 2026-11-16 21:15:54 +00:00 by HumanBean · 1 comment
Owner

What happens

With AMD FSR 3.1 frame generation (#1103), a character's limbs that swing between two game frames (Link's arms, legs and hair while strafing left and right, most visible at 2x) show in the made frames as faded, see-through copies, often where the limb was in the frame before. The game's own frames are clean.

Seen on an RTX 3090 (Linux, by hand at about 37 game fps, and scripted) and an RX 580 (Linux, the AMD lab, scripted with RX_VK_SHOTS + RX_VK_SHOT_MADE=1, each made frame beside the real frame after it).

What is known

  • Not the motion vectors. The game's motion handed to FSR (the temporal chain's, with the skinned draws' own: actor motion) was checked against block matching between the real frames: median error 1.3 px over 3,638 moving points, Link 0.9 px against ~38 px of motion. RX_VK_DLSS_ACTORS=0 (camera motion alone) and RX_VK_FSR_FG_HUD=0 (AMD's HUD handling) double the limbs the same way.
  • It is FSR's interpolation. Forcing the game's motion alone or the optical flow alone changes where the copy lands, not whether there is one. AMD's pass blends the two frames' samples and the optical flow's, and inpaints what neither side is taken to see; a thin limb that moved further than its own width ends up as a partial copy at each place.
  • What didn't help: taking the nearer frame's sample wherever the two frames' samples disagree by more than a threshold (0.2–0.4, or 0.05–0.15), with or without skipping the inpainting there. Made frames on both GPUs still showed the copies, about as AMD's own blend does (tried in #1103, taken out of it).

Suggested approach

Hand FSR's interpolation a mask of the pixels the skinned draws cover (actor motion already draws them, in the chain's motion pass), and on those pixels take only the nearer real frame, warped by the character's own motion: no optical flow, no blend, no inpainting. Characters would step at the game's frame rate instead of ghosting; the rest of the picture keeps FSR's interpolation. Check it the way the RX 580 runs did: made frames beside real frames while strafing past Saria in Kokiri, against the same build without the mask.

Workaround

Frame generation by: Built-in FG (more ghosting overall), or a higher game frame rate (less motion between frames, smaller copies).

## What happens With AMD FSR 3.1 frame generation ([#1103](https://git.magicbean.dev/MagicBeanDev/Longshot/pulls/1103)), a character's limbs that swing between two game frames (Link's arms, legs and hair while strafing left and right, most visible at 2x) show in the made frames as faded, see-through copies, often where the limb was in the frame before. The game's own frames are clean. Seen on an RTX 3090 (Linux, by hand at about 37 game fps, and scripted) and an RX 580 (Linux, the AMD lab, scripted with `RX_VK_SHOTS` + `RX_VK_SHOT_MADE=1`, each made frame beside the real frame after it). ## What is known - **Not the motion vectors.** The game's motion handed to FSR (the temporal chain's, with the skinned draws' own: actor motion) was checked against block matching between the real frames: median error 1.3 px over 3,638 moving points, Link 0.9 px against ~38 px of motion. `RX_VK_DLSS_ACTORS=0` (camera motion alone) and `RX_VK_FSR_FG_HUD=0` (AMD's HUD handling) double the limbs the same way. - **It is FSR's interpolation.** Forcing the game's motion alone or the optical flow alone changes where the copy lands, not whether there is one. AMD's pass blends the two frames' samples and the optical flow's, and inpaints what neither side is taken to see; a thin limb that moved further than its own width ends up as a partial copy at each place. - **What didn't help:** taking the nearer frame's sample wherever the two frames' samples disagree by more than a threshold (0.2–0.4, or 0.05–0.15), with or without skipping the inpainting there. Made frames on both GPUs still showed the copies, about as AMD's own blend does (tried in #1103, taken out of it). ## Suggested approach Hand FSR's interpolation a mask of the pixels the skinned draws cover (actor motion already draws them, in the chain's motion pass), and on those pixels take only the nearer real frame, warped by the character's own motion: no optical flow, no blend, no inpainting. Characters would step at the game's frame rate instead of ghosting; the rest of the picture keeps FSR's interpolation. Check it the way the RX 580 runs did: made frames beside real frames while strafing past Saria in Kokiri, against the same build without the mask. ## Workaround Frame generation by: Built-in FG (more ghosting overall), or a higher game frame rate (less motion between frames, smaller copies).
Collaborator

Possibly related, for context:

  • #929 Game: Frame generation flickers with frame pacing enabled — Related frame-generation artifact report, but it concerns flicker/stutter with different conditions.
  • #1075 Game: DLSS Frame Generation causes artifacts in HUD — Related frame-generation visual artifact, but limited to NVIDIA DLSS HUD artifacts.
Possibly related, for context: - #929 Game: Frame generation flickers with frame pacing enabled — Related frame-generation artifact report, but it concerns flicker/stutter with different conditions. - #1075 Game: DLSS Frame Generation causes artifacts in HUD — Related frame-generation visual artifact, but limited to NVIDIA DLSS HUD artifacts.
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#1236
No description provided.