Guide 02 Features and Settings

Documentation snapshot: November 5, 2026. Community-reported guidance; GPU drivers, capture card firmware and Veyra versions may have changed. See Home for current links.

Veyra User Guide — 02 — Features and Settings in Detail

---

Veyra User Guide - 02 - Features and Settings in Detail

Compiled from Q&A by the author (Likely7) and experienced group members, covering versions 1.4.x to 2.0.x

1. Processing Pipeline Overview

Console/PC ──HDMI──> capture card ──USB (or PCIe)──> Veyra │ NR (DLSS5 neural rendering) → upscaling (line doubling) → frame generation (interpolation) → output window/fullscreen Audio chain: capture card audio → audio compensation/sync → output Player chain: local video file → can be run through NR + upscaling + frame generation too Streaming chain: PS5 (chiaki NG core) or dual-PC streaming → same as above

Core principles (important):

• The "latency" shown in the app = processing time, not end-to-end latency; "most of the latency still comes from the capture card" (the author)

• "Frame generation is the icing on the cake, not a rescue": a 30 fps source already carries 33 ms of render latency, and frame generation cannot change the baseline responsiveness

• The three levers for low latency: capture at 60 (not 30) → NV12 format → PCIe or 10G capture card

2. NR (Neural Rendering / DLSS5) Settings

[What is NR] "nr is dlss5"; "it's just an ai filter" — DLSS5 is used to enhance the picture with AI [NR render resolution] Native / 1440P / 1080P / 900P / 720P / 480P

• Native: the sharpest but extremely demanding; "native nr is too demanding and quickly eats into the compute needed for frame generation" (the author)

• Recommended: use 1080P or 1440P for gaming; "stacking multiple layers looks weird in many games and is only good for screenshots"

• Measured on a 5060 Ti: 1080P rendering uses about half the GPU, 1440P just saturates it

• When using NR, it's best to lock the console/capture input to 1080: "if you turn on nr, locking to 1080 is recommended" (Snicer) [NR layers] The app supports layer stacking (up to 4 layers)

• NR as shipped in 1.4.4: one layer ≈ the effect of two layers in 2.0

• Layer guidance (from 5060 Ti 16G experience): 3 layers at 30 fps, only 1 at 60 fps; "the higher the frame rate, the higher the dlss5 cost"

• Measured on a 5090: single-layer native NR at 8K = 110 load; dual-layer native NR at 4K is usable; "the performance cost is too high for a little extra sharpness — not worth it" [NR exclusion mask] "It's for the areas where you don't want DLSS, like where a game has UI — box them off and exclude them. The feathering around the edges softens the transition" (the author) [People-related parameters] (2.0 color/enhancement page)

• People structure strength: at 30 it doesn't look fake, above 50 it looks "dark and old"; to fix the aged look, "turn the skin structure strength down a bit"

• People mask: keep it under 40, "otherwise it really does look AI-generated"; don't push the face effect to 2, "too AI" [Overall strength] "just set it to 2", "pull it to 1.5-2"; "raise it if you want DLSS5 to step in more"; "lower the overall change strength and the color tone will be close to native" [Conditions for DLSS5 to take effect] (group member experience): the person has to be large enough (a face of at least 80x80 clear pixels) to be recognized; stacking multiple layers makes the effect more obvious; colors may shift when the camera turns (this is normal)

3. Upscaling (Super Resolution) Settings

[Options] Target 4K/6K/8K + quality tier (low/medium/high/highest, etc.) [Key points]

• "The difference between upscaling settings isn't big" (the author); measured: "there's no real difference between 4k and 8k images, only more latency"

• "Capturing 4K and then upscaling to 4K makes no perceptible difference"; on a 4K screen, capturing 1080P and upscaling "will definitely still be blurry"

• Upscaling is demanding: "set upscaling too high and it can't keep up"; "with medium upscaling, power draw drops a lot"

• If you see flicker, switch video upscaling to the medium tier to mitigate it

• Upscaling pays off most on low resolutions (1080P/2K); recommended: "capture 4K → output 1080 → upscale to 4K" (the zL approach)

• 2026-10-03 the author's advice (Ghost of Yotei comparison): "4k upscaling + native NR is clearer than 8k upscaling + 1440 NR" [Order] NR must come first ("nr must come first" — Snicer); low latency mode swaps the processing order of "DLSS5 and upscaling"

4. Frame Generation (Interpolation) Settings

[Multiplier] 2x/3x/4x/dynamic frame (6x supported on the 30/40/50 series since 1.4.0; experimental on AMD cards) [Engine selection] Available in the app: XESS / DLSS / AMD optical flow / NVIDIA optical flow / zero motion

• Group consensus: "xess frame generation is widely considered good" (less jelly); "DLSS frame generation has the worst image errors (including ghosting)"

• "dlss 5x and 6x are basically unusable" (Regulus); most people use XESS 2x~4x

• NVIDIA App AI frame insertion (40/50 series): very low GPU usage (single-digit %), and it can stack with the app for even higher frame rates "AMD optical flow + XESS 2x + NV App AI frame generation = 30 fps close to 120 fps" (the author's 1.4.0 announcement) [Optical flow options] (motion estimation)

• "Don't use dis optical flow, it's garbage" (the author); recommendation: "use amd optical flow with xess"

• For image flicker, check this first: "many flicker issues come from optical flow not being enabled in NR rendering"

• NVIDIA optical flow: least flicker, slightly more demanding, good results; AMD optical flow: lower usage, slightly worse and may flicker

• "optical-flow frame generation costs 3-4ms more than zero-motion frame generation"

• After 2.0, "in the next version XESS optical flow can be turned off" (the author) [Content cadence (60→30)]

• When to use: when the game actually runs at 30 fps but the capture is 60 fps (duplicate frames), choose "capture 60→30"

• "If a game runs at 30fps and you capture at 60fps, half the frames are duplicates — turn on 2x frame generation and it's unplayable" (Kallur)

• Recommended recipe for 30 fps PS5 games: game on quality/30 fps mode → capture 4K30 or 2K60 → content cadence 60→30 → NR (1080/1440) → upscale 4K → XESS 2x (verified by several group members) [Measured latency reference] "with DLSS5 on plus 4x frame generation, latency is 30~40ms" (the author); "1.4.1 includes a lot of fixes… the old pipeline was a bit congested, so frames were generated but missed their display slot — that's been optimized"

5. Other Options on the Quality Page

• 1080P internal processing: lowers DLSS5 cost and noticeably reduces power draw, "with basically no impact on the picture"; only takes effect above 1080P; "if the source is already 1080P and upscaling is off, it makes no difference whether it's on or off" (the author)

• Low latency mode: swaps the processing order of DLSS5 and upscaling (DLSS5 first, then upscaling → possible tearing); the switch has no effect when only DLSS5 is on; "if upscaling is off, it makes no difference either way" (the author)

• Output stabilizer anti-flicker: "it's supposed to prevent flicker, but it doesn't seem to do much" (the author); in early 2.0 this switch could freeze the image (fixed in 2.0.2)

• Temporal anti-flicker: same as above (the author's verdict: doesn't do much)

• Streaming compatibility mode (experimental): "doesn't do much" (the author); if OBS recording stutters, the author suggests turning on "obs compatible" (in settings)

• Feathering: softens the transition at the edge of the exclusion mask

• Auto mask / UI correction: used to protect UI areas (no detailed explanation in the group)

• Capture card flip: for when the picture is flipped in RGB format (1.3.2+)

• Color page (1.4.0+): modeled on LR/PS, with exposure/contrast/highlights/shadows/whites/blacks/clarity/texture/skin structure strength; LUT import is supported; "if the picture has a color cast, try the in-app tint adjustment"

• Screenshots: what's saved is the processed image — "taking a screenshot with 8K upscaling on saves a 40MB image"

6. Input and Audio Settings

[Capture format mapping] SDR → YUY2 / NV12 / RGB24 HDR → P010 (+ manually set the input color to PQ HDR10) "sdr yuy / hdr p010" (Yoloッ's pairing table) Native format ranking (bandwidth/quality): RGB24 > YUY2 > NV12 > MJPEG Latency ranking (low→high): NV12 ≈ YUY2 < RGB24 < MJPEG P010: 10bit 4:2:0, carries HDR data ("p010 is itself an SDR format but carries HDR data" — 咖喱) [Input color] Auto detect or manual (PQ HDR10 / HLG); "go manual, auto has issues" (the author); choosing the wrong one makes it "output with SDR colors, and the colors look strange" [Audio] Select the device with the same name as the capture card; "latency optimization" can be turned off; "monitoring" must be on to hear the audio on the PC; professional mode has "audio compensation"; A/V sync offers auto/manual, and manual is steadier [Dolby] Dolby Atmos is supported from 1.4.0 (together with an audio/video splitter setup)

7. Performance Tiers and Machine Reference (measured data from the group)

• 4060 laptop: in enhanced mode the GPU is maxed out, 60 in / 45 out runs smoothly; 3060 laptop at 60 in / 20 out stutters noticeably

• 4060 8G: NR only works at 1080P, one layer already blows past VRAM → drop to 900P; with OBS streaming on, 1080P flickers → 900P

• 5060 Ti 16G: 1080P rendering uses half the GPU, 1440P saturates it; "4K60 NV12 + 2K NR rendering is easy (but no frame generation)"

• 4070 Ti: can only handle 1080 capture plus 2x dlss frame generation (older versions); 4070S owners should watch out for "compatibility with NR files" (the 40 series has a compatibility issue where "the GPU hangs after a while" — restart the app; NR input at 1080P is more stable)

• 5070 Ti: true native 4K60 can't be saturated; 1440P rendering is recommended

• 5080: DLSS5 in fullscreen can drop into the 40s → pick a lower input resolution and then upscale

• 5090/5090D: only 8K upscaling plus dual-layer NR really pushes it; power draw is 450-600W, so capping power or undervolting is recommended

• 30 series: use XESS frame generation + AMD optical flow; "a 3060 probably can't handle" 6x

• 20 series: frame generation is not recommended ("don't try it on a 20" "it stutters like crazy even if you turn it on"), only try low XESS multipliers

• AMD cards (9000 series): DLSS5 is an experimental project; for video playback use the community build of 大力喜鹊 (lmxxf/ENTROPY repo)

8. Recommended Configuration Templates (by scenario)

[Scenario 1: PS5 30 fps games (quality mode) — the most comfortable] Console outputs 4K60 → capture card captures 4K30/2K60 HDR or SDR → content cadence 60→30 → NR 1080P or 1440P → upscale 4K (medium-high tier) → XESS 2x → optical flow AMD/NVIDIA depending on the machine Group member measurement: "capture 2k 60 hdr, game in performance mode, upscale to 4k, turn on 2x frame generation — picture quality and latency are both comfortable" (Wwz) [Scenario 2: PS5 60 fps games] Capture 2K60 or 4K60 (SDR NV12) → NR 1080P/1440P → upscale 4K → 2x frame generation → 120Hz monitor [Scenario 3: pure low-latency competitive play (quality aside)] Capture 4K30 NV12 or 2K60 → only XESS 2x frame generation (no NR) → lowest latency "frame generation alone keeps latency perfectly acceptable"; low latency mode + vsync as needed [Scenario 4: Switch 2] 2K60 capture (G50/27P is enough) → NR 1080P → upscale 2K/4K → 2x frame generation; Zelda and similar games can reach 120 NS2 capture note: a card without HDR support makes the NS2 turn HDR off automatically and the picture gets darker — either turn HDR off in the settings or switch to a P010 card [Scenario 5: dual machine (one for gaming, one for processing)] Gaming PC (high resolution, high refresh) → capture card → processing PC with Veyra maxed out → output; or 2.0 dual-machine streaming (500M bitrate, about 16ms at 4K) "the 9070 runs the game with one layer of dlss5 locked at 60, then feeds the capture to the 5060 Ti for another layer" (牛腩盖饭) [Scenario 6: using it as a player] Local video → NR + upscaling + frame generation all on → "video playback is so complete that I don't even bother watching Dolby Vision" (a user) Exporting 8K VR video has its own workflow (import → DLSS → export → SteamVR)

9. Shortcuts and Tips (collected from the group)

• In fullscreen, the left/right arrow keys skip forward/back 10s (1.4.4+); in fullscreen you can summon/hide the panel

• Desktop/windowed mode: dragging the window's "frosted glass" used to eat CPU

• Frame rate stabilization tip (Yoloッ): "add Veyra in RTSS, choose NVIDIA Reflex as the frame limiting method, and lock to your post-frame-generation frame rate"; RTSS latency ranking: Reflex < async < front edge sync

• Raising priority: "use a scheduling tool to keep the app on the big cores only and set the CPU priority to high"

• Checking for capture card contention: OBS, PotPlayer, and streaming companion apps can all occupy it

• Running multiple versions side by side: just use different folders

• Presets: can be exported/imported and shared; group member presets such as "Ghost of Yotei v2.vpreset" circulate

• Log location: the logs folder in the app's root directory (always attach it when reporting a bug); crashes produce veyra-crash-*.dmp

10. Limitations and Known Gaps (so you don't waste time)

• Streaming: officially the PS5 is limited to 1080P; an XSX streaming mode doesn't exist (PS5 only); jailbroken consoles can't stream

• Local window capture (1.4.2+): "you probably can't game with it, but watching a video works"; there's a ban risk, so don't use it for online games

• Recording: the app has no recording module ("for recording just capture with obs — use game capture, which has no ui")

• Custom frame rates: 40 fps/custom can't be frame-generated (it drops frames, a "134678" kind of uneven rhythm)

• Real-time translated subtitles: no plans yet; phone screen casting: no solution yet

• AMD cards: NR is limited (experimental); macOS version: porting would be difficult

• VRR: once captured it can be treated as unsupported; it's only useful on passthrough; a fixed 30 or 60 fps gives the best experience

---

12. This Update: New Features in 2.0.2 → 2.1.2 (added 2026-11-05)

The original text above covers up to 2.0.x; what follows was added later and is written in menu order.

1. NVIDIA Low Latency: Off / On / On + Enhanced

• It uses NVAPI Reflex; the list and the node share the same control, and the old "low queue" setting migrates automatically.
• Under high load choose "On", not "On + Enhanced": Enhanced makes the driver wait an extra 4–18ms per frame, and on-site RTX5090 testing showed 8–21% fewer generated frames (the author's 11-02/11-03 notes). 2.1.2 adds a hint in the UI.

2. Dual Upscaling (base upscale → shrink → second upscale)

• Only supported with RTX video upscaling, off by default; details look more solid, but VRAM and compute usage double.
• Mutually exclusive with "shrink after upscale"; the two can't be on at the same time.

3. Shrink After Upscale (supersampling)

• A single upscale to a larger size, then a precise shrink back in linear light; NR/frame generation run at the post-shrink size.
• Pick this to balance sharpness and frame rate; only 5090-class machines should try dual upscaling.

4. Frame Generation Latency Budget

• Options: lowest default / auto / +1 / +2 source frames / 0–100ms; the limited output pool is allocated according to the budget.
• Reserves an extra 1.3–2.3GB of VRAM; if there isn't enough it declines and restores automatically, so it won't lock you up.
• Does not apply to file playback, images, export, or XeSS. If it's declined, lower upscaling or turn off dual upscaling.

5. Vsync Wait (3ms cap)

• With explicit vsync and output ≤80fps, the wait for generated-frame capacity grows with callback jitter, capped at 3ms.
• Result: generated-frame submission gaps go from 37.7/min → 5.6/min (short sample).

6. Automatic Duplicate Frame Removal

• Whole-frame deduplication that tolerates out-of-range components of about five parts per million; fixes the wrong and missed deletions of older versions (2.1.2).

7. RX7000 (RDNA3) NR

• Since 2.1.2 the device is selected by LUID, so multi-GPU setups (discrete + integrated) no longer go black; still experimental.

8. Other

• 60→30 content cadence keeps the first copy; F7/F8 shortcuts; Dolby/DTS carrier; fixed volume interaction in minimal mode.

13. This Update: New Group Measurements (with dates)

• After the update there are two extra "progressive" specifications; the author: "pick the first one" (the author 10-07).
• Shortcuts: "hold V for a quick before/after comparison; in fullscreen, Home brings up the adjustment menu" (the author 10-12).
• If you only want the clean picture: "turn on OBS compatibility mode in the settings, then choose game capture in OBS" (the author 10-10); UI glitches from enabling OBS compatibility are normal (the author 10-07).
• Installing the wrong package causes problems: "there are two packages now, one for AMD and one for NVIDIA" (the author 10-10).
• Reducing facial shine: raise the NR internal processing resolution (Minions 10-07: "the best way is to increase the nr resolution — native is best"), then lower the light/dark variation slightly; the tradeoff is "the less shine on the face, the smaller the gain".
• One NR layer now comes close to multiple layers in older versions: "before, NR×3 → great picture, dead GPU; now NR×1 + picture tuning → a similar result" (四夕 10-06).
• Frame generation from a 30 fps source always comes up short: "if you want frame generation, aim for at least 60 fps" (Snicer 9070xt 10-07); a ready-made recipe: "Ghost of Yotei at 30 fps, VFG high quality, 3x" (the author 10-08).
• If frame generation underperforms, check two things first: whether 60→30 was enabled by mistake; and set the capture frame rate to 59.94 (the author 10-12).
• AMD cards can't enable VFG: "no, that's NVIDIA's tech" (the author 10-07).
• Latency accounting: one NR layer is about 3.x ms on a 5090 and about 10ms on a 5060 (继续吹 10-12), roughly +1.7ms per layer (拔筷四顾心茫然 10-22).

Chapter 02 Addendum: New Settings Q&A (10-06 to 11-05)

Filtered from the records of 7 chat groups, 2026-10-06 to 11-05, 2965 relevant messages in total.

Highlights of This Round

• 2.1.x adds 7 items: NVIDIA low latency with three levels, dual upscaling, shrink after upscale, frame generation latency budget, vsync wait (≤3ms), automatic duplicate frame removal, and RX7000 NR.
• Three low latency levels: off / on / on + enhanced; under high load choose "On", because Enhanced makes the driver wait an extra 4-18ms per frame and produces 8-21% fewer generated frames (the author 11-02).
• Dual upscaling vs shrink after upscale: pick one, they can't both be on; ordinary setups should choose "shrink", and only 5090-class machines should try dual.
• Frame generation latency budget: reserves 1.3-2.3GB more VRAM, and declines with automatic restore if there isn't enough; if it's declined, lower upscaling or turn off dual.
• The update adds a "progressive" option: the author says "pick the first one" (the author 10-07).
• Shortcuts: "hold V for a quick before/after comparison; in fullscreen, Home brings up the menu" (the author 10-12).
• Clean picture only: "turn on OBS compatibility mode in the settings, then choose game capture in OBS" (the author 10-10); UI glitches from enabling compatibility are normal.
• Installing the wrong package causes problems: "there are two packages now, one AMD and one NVIDIA" (the author 10-10).
• NR for less facial shine: raise the NR internal processing resolution (Minions 10-07), then lower the light/dark variation; the less shine, the smaller the gain.
• One layer now approaches multiple layers: "before, NR×3 killed the GPU; now NR×1 + picture tuning is close" (四夕 10-06).
• Frame generation from 30 fps has a ceiling: "for frame generation, at least 60 fps is recommended" (Snicer 9070xt 10-07); ready-made recipe "Ghost of Yotei at 30 fps, VFG high quality, 3x" (the author 10-08).
• Latency accounting: one NR layer is about 3.x ms on a 5090 and about 10ms on a 5060 (继续吹 10-12), roughly +1.7ms per layer (拔筷四顾心茫然 10-22).

Group Members' Own Words (with dates)

"@Likely7 the issue where the picture freezes after turning on frame generation on xsx streaming hasn't shown up so far — everything is normal except the NR black screen"

— Snicer 9070xt (10-06)

"I'm planning to drop to 2k60 capture and then upscale with veyra [旺柴], since ps5 games aren't native 4k anyway the difference shouldn't be much, but it buys lower latency"

— Focus (10-06)

"I found upscaling latency is quite high — 2k upscaled to 4k is about 15ms of latency"

— Minions (10-06)

"So for gta6 at 30 fps, how much end-to-end latency does capture plus frame generation roughly add"

— 无敌Siuuu.... (10-06)

"As long as what you capture is the top-end picture your capture card can handle, there's no latency — for example if your game is 30 fps and you turn on ai frame generation, the picture has slight latency"

— Sadjoy (10-06)

"Even pc games have slight latency once ai frame generation is on"

— Sadjoy (10-06)

"I don't have a 5090 either, no idea whether upscaling latency would be lower on a 5090"

— Minions (10-06)

"CR latency can hit 30 milliseconds while having both frame generation and 4k upscaling"

— 威猛先生 (10-06)

---