Frame generation is the most misunderstood number in a modern performance overlay. The counter doubles or quadruples, but the frames padding that number are interpolated between rendered ones — they carry no fresh input state, and the extra stage they add to the render pipeline can push latency up even as the FPSframerate. How many images (frames) the game shows per second; higher = smoother motion. 60 fps is a common target. readout climbs. That trade-off is precisely why NVIDIA ships Reflex inside the DLSS 3 stack and why AMD wires Anti-Lag 2 into FSR frame generation. This guide walks the pipeline from render queue to scan-out using both vendors’ own documentation, then draws the practical lines: when the pairing rescues the experience, and when you should leave frame generation off.
What this covers: how frame generation interacts with the render queue, why Reflex and Anti-Lag 2 matter when generated frames are on, and how DLSS 3/4 and FSR 3.1 differ on latency and frame pacing. Why it matters: a 200 FPS counter built from interpolated frames can feel worse than 90 rendered FPS when the queue is deep — and the fix is a specific latency knob, not more generated frames. Who should care: anyone tuning a modern GPUGPU. Graphics Processing Unit — the chip that renders the game's visuals; the main driver of framerate and image quality., deciding whether to enable frame generation in a competitive shooter, or wondering why a “faster” game suddenly feels floaty.
The render queue is where latency hides
Every rendered frame passes through a queue before the GPU touches it. NVIDIA’s latency walkthrough describes the render queue as “designed to keep the GPU constantly fed by always having work buffered for the GPU to do,” which maximizes throughput — and, in NVIDIA’s words, “can introduce latency.” When the CPU finishes frames faster than the GPU consumes them, submitted work piles up; NVIDIA’s pipeline example shows a GPU-bound case where the CPU runs two frames ahead. Every queued frame is a frame your mouse click waits behind before it reaches the display.
AMD describes the same mechanism from the driver seat. Radeon Anti-Lag, the original driver-level feature, “introduces a carefully calculated delay into the driver-side processing of game’s graphics commands, such that CPU and GPU frames are optimally aligned,” preventing CPU frames from running too far ahead of GPU frames. AMD’s Anti-Lag 2 integration guide makes the queue depth visible: with the latency monitor’s bypass key held, native latency reads “close to an integer number 3, 4 or 5” GPU frames — that integer is your queue, counted.
FPS is a throughput metric; latency is a delay metric. A deep queue maximizes the first and inflates the second, which is why both vendors sell a tool whose entire job is keeping the queue shallow.
What frame generation adds to the pipeline
Frame generation does not render more frames — it inserts predicted ones. AMD’s FSR 3 documentation states it plainly: the interpolation “inserts 1 frame between existing ones which can therefore enable up to 2x the framerate in supported games.” DLSS 3, per its announcement, feeds current and prior game frames, an optical flow field from the RTX 40 Series’ Ada Optical Flow Accelerator, and game engine data such as motion vectors and depth into a convolutional autoencoder that generates one additional frame for each rendered frame. In both cases the new frames sample no input of their own; your clicks and mouse movement still land only on rendered frames.
The interpolation stage also costs real time inside the frame budget. NVIDIA’s DLSS 4 engineering write-up reports the DLSS 3 frame generation model produced a single 4K frame in around 3.25 ms on a GeForce RTX 4090, while DLSS 4 Multi Frame Generation generates each of its three additional frames in around 1 ms on average on an RTX 5090. AMD’s FSR Frame Generation page lists roughly 2.2 ms per generated 4K frame on a Radeon RX 9070 XT. Faster models shrink the cost; none of them remove the stage.
That stage is measurable, and both vendors instrument it. NVIDIA’s Reflex Test Utility report breaks PC latency into pipeline stages including “R>FG — DLSS FG start to GPU end,” and its sample report is literally captioned “Reflex ON, Frame Gen x2” — frame generation is a line item in the latency budget, not a free multiplier. AMD’s equivalent is explicit in the Anti-Lag 2 SDK guide: Anti-Lag 2 “requires some special attention when AMD FSR Frame Generation is enabled,” needing extra calls so the SDK knows “whether the presented frames are interpolated or not.”
AMD states the design consequence without hedging: “Interpolating a frame adds latency – this is by design.” The whole reason each vendor bolts a latency stack onto its frame generator is to buy that latency back. For the display-side view of why generated frames can look smooth while feeling behind the crosshair, our existing explainer on Frame Generation’s Hidden Latency and the Reflex Offset covers the measurement side.
NVIDIA’s pairing: Reflex rides inside DLSS
NVIDIA never shipped frame generation alone. DLSS 3 was announced in September 2022 as three technologies together — Super Resolution, Frame Generation, and Reflex — and NVIDIA’s claim was about the stack, not the interpolation: “DLSS 3 also incorporates NVIDIA Reflex, which synchronizes the GPU and CPU, ensuring optimum responsiveness and low system latency,” and “When compared to native, DLSS 3 can reduce latency by up to 2X.”
What Reflex actually does is queue surgery. The Reflex SDK documentation describes its low latency mode as aligning “game engine work to complete just-in-time for rendering, eliminating the GPU render queue and reducing CPU back pressure in GPU-bound scenarios.” Because the CPU starts its work later, inputs can be “sampled at the last possible millisecond.” NVIDIA’s 2020 deep dive calls the result a dynamic framerate limiter that keeps the pipeline in the latency sweet spot, and it runs far down the stack: GeForce GTX 900 Series and newer (driver 456.38 or later). By January 2025, NVIDIA reported Reflex was integrated in over 100 games, reduces PC latency in top competitive games by an average of 50 percent, and that over 90 percent of gamers leave it on.
Reflex 2, announced at CES 2025, adds Frame Warp: as a frame finishes rendering, the CPU calculates the newest camera position from the latest mouse input, and the GPU warps the finished frame to it “just before the rendered frame is sent to the display.” NVIDIA’s numbers: up to 75 percent PC latency reduction; in THE FINALS at 4K max settings on an RTX 5070, latency falls from 56 ms to 27 ms with Reflex Low Latency and to 14 ms with Reflex 2; in VALORANT at 800+ FPS on an RTX 5090, PC latency averages under 3 ms. Frame Warp debuts on GeForce RTX 50 Series GPUs first, with support for other GeForce RTX cards to follow in a future update.
The pairing is enforced, not suggested: NVIDIA’s Streamline Reflex QA checklist requires the Reflex test report to pass “without warning with DLSS Frame Generation” enabled — a Reflex integration is tested with frame generation on, because that is the configuration it must survive.
AMD’s pairing: Anti-Lag 2 beside FSR frame generation
AMD arrived at the same architecture by a rougher road. Driver-level Anti-Lag, the original feature, runs on Radeon RX 400 Series cards and newer. In 2023, AMD paired FSR 3 with Anti-Lag+, a driver-injected per-game variant — which tripped anti-cheat systems in titles including Counter-Strike 2, and AMD withdrew it through a driver update (as reported by VideoCardz at the time). Anti-Lag 2, launched in May 2024, is the correction: an in-game SDK developed with Valve that requires developer integration.
The mechanism mirrors Reflex’s, described in AMD’s own terms. Anti-Lag 2 keeps the driver-side alignment of Anti-Lag 1 but moves “the point of insertion of the delay… at the optimal point inside the game’s logic, just before the user controls (mouse/gamepad/keyboard) are sampled,” which AMD says achieves “a significantly greater latency reduction.” Hardware floor: RDNA 1 — Radeon RX 5000 Series and newer — with Adrenalin 24.6.1 for DirectX 11, 24.7.1 for DirectX 12, and 24.9.1 for Vulkan. AMD’s technical preview release notes measured “a decrease in latency in Counter-Strike 2 of up to an average of 37%” across RX 7900 GRE, RX 7800 XT, RX 7600 XT, RX 7700S and the Ryzen 7 8700G’s Radeon 780M, crediting “in-game pacing of the CPU with the GPU, particularly in GPU-intensive scenarios.”
The FSR side of the pairing evolved with it. At FSR 3’s launch in 2023, AMD recommended Anti-Lag+ on RX 7000 Series cards to “further lower the latency while using FSR 3,” with driver Anti-Lag covering RX 6000 and below. After Anti-Lag+‘s retirement, the FSR 3.1.1 Frame Interpolation Swapchain (October 2024) added Anti-Lag 2 support directly, and the current SDK handles the interpolated-frame bookkeeping from FSR 3.1.1 onward. AMD’s own enablement steps on its FSR page close the loop: “We also recommend enabling Radeon Anti-Lag 2 in games that support it.”
The two latency stacks, side by side
| Latency technology | Layer | What it does | Hardware floor | Frame generation pairing |
|---|---|---|---|---|
| NVIDIA Reflex Low Latency | In-game SDK (Streamline/NVAPI) | Empties the GPU render queue; samples input just-in-time | GeForce GTX 900 Series+ | Ships inside DLSS 3; QA-tested with DLSS FG enabled |
| NVIDIA Reflex 2 Frame Warp | In-game SDK | Warps the rendered frame to the newest camera position before scan-out | GeForce RTX 50 Series (at debut) | Up to 75% latency reduction; 56 ms to 14 ms in THE FINALS |
| AMD Radeon Anti-Lag | Driver | Inserts a delay in driver-side command processing to align CPU and GPU | Radeon RX 400 Series+ | The fallback AMD recommends for FSR 3 on RX 6000 and below |
| AMD Radeon Anti-Lag 2 | In-game SDK | Inserts the delay inside game logic, just before input sampling | Radeon RX 5000 Series (RDNA 1)+ | FSR 3.1.1 swapchain support; tracks which frames are interpolated |
Note what the table does not say: neither Reflex nor Anti-Lag 2 is a frame-generation feature. Both are queue-draining tools that happen to be the mandatory companions for it — Reflex on the NVIDIA side, Anti-Lag 2 on the AMD side. FSR 3 itself is open, and AMD’s own documentation lists GeForce GTX 10 Series and above as supported hardware for it, but the latency stack that pairs with it is Radeon-only.
The numbers both vendors publish
The two vendors’ headline latency claims sound contradictory and are not. AMD’s FSR 3 technical article states that “the latency induced from enabling both AMD FSR 3 upscaling and frame generation is typically less than the native rendering experience without AMD FSR 3 enabled.” NVIDIA’s DLSS 4 announcement reports that in a Cyberpunk 2077 scene on RTX 5090, the full DLSS 4 stack multiplies performance by over 8X “and PC latency is halved” versus brute-force rendering; its research write-up shows Avowed going from 35 FPS to 215 FPS with input latency falling from 114 ms to 43 ms.
Both claims compare a full stack — upscaling raising the rendered base rate, plus latency tech draining the queue — against native rendering with none of it. Neither vendor claims interpolation alone reduces latency. AMD’s own launch measurements show the interpolation cost directly:
| Game, GPU, resolution (AMD launch testing) | Native (FSR 3 off) | FSR 3 upscaling only | Upscaling + frame generation |
|---|---|---|---|
| Forspoken, RX 7900 XTX, 4K | 81 ms (55 FPS) | 48 ms (90 FPS) | 55 ms (164 FPS) |
| Forspoken, RX 6800 XT, 1440p | 81 ms (56 FPS) | 55 ms (83 FPS) | 58 ms (160 FPS) |
| Immortals of Aveum, RX 7900 XTX, 4K | 72 ms (57 FPS) | 37 ms (107 FPS) | 40 ms (167 FPS) |
| Immortals of Aveum, RX 6800 XT, 1440p | 63 ms (68 FPS) | 37 ms (113 FPS) | 43 ms (128 FPS) |
Read the middle two columns: adding frame generation to upscaling costs 7 ms in Forspoken at 4K and 3 ms in Immortals, while the FPS counter roughly doubles. The stack still beats native-off — upscaling did the heavy lifting — but frame generation itself added latency on top of the upscaled baseline. That is the trade in one table, and it is AMD’s own data.
Pacing is the other half of the trade
Latency measures delay; pacing measures regularity, and uneven delivery reads as stutter even at high average FPS. The vendors moved pacing hardware-side at different speeds. DLSS 3 Frame Generation “used CPU-based pacing with variability that can compound with additional frames, leading to less consistent frame pacing between each frame,” in NVIDIA’s own retrospective — which is why DLSS 4 on the Blackwell architecture shifts the logic: “hardware Flip Metering… shifts the frame pacing logic to the display engine, enabling the GPU to more precisely manage display timing.” Our DLSS 4 Multi-Frame Generation: Latency and Reflex guide covers what that means for RTX 50 owners in practice.
AMD’s FSR frame generation paces through its replacement swapchain, and AMD documents the expected warts: with VSync enabled the present-to-present timing graph shows a “zig-zag” pattern, which AMD calls expected behavior, recommending a game-side frame limiter at half the refresh rate for the most consistent timing. The current FSR Frame Generation page adds a stability warning worth remembering in the field: the feature “requires unencumbered access to the swap chain for best frame pacing results,” and third-party software that intercepts DXGI calls — capture tools, overlays — can cause pacing issues.
When to leave frame generation off
AMD publishes a floor, NVIDIA does not, and the difference is itself informative. AMD’s FSR Frame Generation documentation states the technology “runs best when interpolating from a minimum of 60 fps pre-interpolation,” that going below 60 fps “is not recommended as interpolation artifacts become more prominent,” and that “sub-30fps pre-interpolation should be absolutely avoided” — the same 60 FPS guidance AMD gave at FSR 3’s launch. NVIDIA, by contrast, publishes no strict numerical minimum for DLSS Frame Generation in its official documentation, as Wccftech’s 2026 setup guide notes; its materials lean on Reflex riding along with the stack instead, and modern titles frequently auto-enable Reflex when DLSS frame generation turns on.
| Your situation | What the vendors publish | The practical call |
|---|---|---|
| Base rate under 30 FPS | AMD: “absolutely avoided” | Fix settings or upscale first; frame generation cannot rescue this |
| Base rate 30–60 FPS | AMD: below 60 “not recommended,” artifacts more prominent | Defensible only in slow, controller-driven games |
| No numerical floor stated | NVIDIA: no published DLSS FG minimum; Reflex pairs with the stack | Judge by feel, not by the counter |
| Competitive shooter, mouse aim | Reflex and Anti-Lag 2 exist for exactly this; generated frames add a stage | Run the latency stack, leave frame generation off |
The competitive row is where the two technologies in this guide’s title converge. Inputs land on rendered frames, so a displayed 240 FPS built on a 60 FPS base still responds like 60 FPS plus interpolation delay — and both vendors position their latency SDKs in esports titles where that difference decides rounds. Wccftech’s guidance is blunt and matches the vendor data: frame generation does nothing to accelerate game simulation, input processing, or the baseline rendering rate, so in competitive play you prioritize native frames plus Reflex or Anti-Lag 2 and skip the interpolation. Frame generation is a finishing tool for an already-responsive game on a high-refresh display — not a rescue for one that is not.