DLSS and FSR Blurry or Ghosting? Upscaling Artefacts, Render Resolution and Quality Modes Explained

2026-10-22 ｜ GPU Repair ｜ About 11 minutes ｜ HK GPU Fix Editorial

Key Takeaways

DLSS, FSR, and XeSS render at a lower resolution and reconstruct the full picture, trading source detail for speed. Blur, ghosting, and flicker on thin detail are usually reconstruction limits, not a failing graphics card. Judge them in motion, choose the highest-quality mode that still meets your frame-rate target, and test with upscaling off before blaming hardware.

Which Problem Is This Article About?

If the picture looks wrong after you turn on DLSS, FSR, or XeSS — soft detail, a faint trail behind moving objects, or fine detail that flickers as the camera moves — and the problem changes when you change the quality mode, this article is for you. If instead the game freezes briefly on first entry into a new area and then smooths out, that is shader compilation stutter, a time-based caching problem, not an upscaling problem. The toggle test separates them: upscaling artefacts move with the upscaler setting, while shader stutter fades with repetition regardless of the upscaler. Neither problem, on its own, is evidence of a failing graphics card.

What Upscaling Artefacts Look Like

After turning on DLSS, FSR, or XeSS, the game runs faster, but the image looks wrong in motion: soft or blurry detail, a faint trail or ghost behind moving objects, fine detail that flickers or shimmers — fences, grilles, foliage, power lines, distant railings — or edges that crawl and sparkle when the camera moves. Crucially, these artefacts are often invisible in a still screenshot and obvious in motion, or the reverse: a still can look sharp while motion smears. Judge upscaling in motion, not from a single still frame.

The complaint pattern is telling: it looked fine before the upscaler was turned on, it gets worse the more aggressive the mode, and it is worst on thin or fine detail and fast movement. That combination points at upscaling reconstruction, not at a dying GPU. Compressed online video also hides fine detail, and still comparisons hide ghosting and shimmer, so online DLSS versus native stills are weak evidence for motion artefacts. Describe and test artefacts in motion terms: what happens when the camera pans, when an object crosses the frame, and when light changes.

Core Mechanism: Render Resolution Versus Output Resolution

Upscaling renders the game at a lower internal, or render, resolution, then reconstructs an image at the monitor's output resolution. The performance gain comes from shading fewer pixels; the image-quality risk comes from having to invent the missing pixels. This is not ordinary stretching. Modern temporal upscalers combine the current low-resolution frame, motion vectors, depth information, and history: reprojected colour data accumulated from previous frames. A neural-network model — DLSS, and FSR 4 on supported hardware — or an analytical temporal algorithm, as in FSR 2 and 3, weighs these inputs to reconstruct edges, sub-pixel detail, and anti-aliasing in one pass.

Motion vectors are per-pixel data from the engine saying where each thing moved since the last frame. Without accurate motion vectors, the upscaler cannot place old detail correctly in the new frame. Depth information helps it tell what is in front of what, and history lets it reuse detail seen in earlier frames.

Jitter is the other key idea. To gather sub-pixel information over time, the engine shifts, or jitters, the camera and sample position slightly each frame. Over several frames the upscaler accumulates different sub-pixel samples of the same surface and combines them into a sharper result. This is why temporal upscalers can sometimes look better than native rendering with weak anti-aliasing: they reduce temporal aliasing and shimmer that native rendering also has. It is also why anything that breaks the history — camera cuts, fast motion, bad motion vectors — breaks the image.

Because old frames can be wrong for the current frame, upscalers clamp or discard history that disagrees with the current frame's neighbourhood, using neighbourhood clamping and disocclusion detection via velocity and depth differences. When this works, ghosting is suppressed. When it fails or is too permissive, old colour bleeds into the new frame as a trail.

A trap writers and readers must avoid is confusing linear percentages with pixel counts. Quality-mode percentages are per dimension, or linear. Quality at 66.7 percent linear means about 44 percent of the pixels, since 0.667 squared is about 0.44. Performance at 50 percent linear means 25 percent of the pixels. Ultra Performance at 33.3 percent linear means about 11 percent of the pixels. Saying Performance only drops to half resolution understates it: it drops to a quarter of the pixel count. Always say per side, or give both numbers.

Quality Modes: The Standard Table

Ultra Performance is described as intended for 8K output in the case of DLSS, or for maximum frame-rate use. At 1080p output its 640 by 360 internal image leaves very little source detail to reconstruct from, so artefacts are expected, not surprising. One secondary source's DLSS 4 table gave different percentages, such as Quality around 90 percent and Performance around 67 percent; that conflicts with the standard preset table above and may reflect dynamic or override behaviour or an error, so it is not used here. Individual games may expose custom resolutions or a dynamic mode: Auto adjusts render resolution to hit a frame-rate target, and the standard table is the preset table, not a promise for every custom or dynamic implementation.

FSR 3's own documentation also lists a sequence length per mode — Quality 18, Balanced 23, Performance 32, Ultra Performance 72 — roughly, how many jittered samples the temporal accumulation uses. Longer sequences at lower render resolutions are part of why aggressive modes can look temporally unstable: more history is being leaned on, from less source detail. Sequence length is not a user-facing setting, so it should be mentioned qualitatively, not presented as something to tune.

Output resolution changes the advice. At 4K output, Performance still starts from 1920 by 1080, a full HD source, so it can look good. At 1080p output, Performance starts from 960 by 540, and Ultra Performance from 640 by 360, so visible artefacts should be expected. This is why Performance looking great at 4K but bad at 1080p is a consistent finding, not a contradiction. At 1440p output, Quality starts from about 1707 by 960, which is 2560 by 1440 scaled by two-thirds per side; use the table rather than inventing equivalents for other modes.

FSR and DLSS Are Not Single Things: Versions Matter

FSR 1 is spatial only: it upscales the current frame with edge-adaptive upsampling and contrast-adaptive sharpening, using no motion vectors and no frame history. It shows more shimmering and aliasing, although the user interface stays sharp because only the 3D scene is scaled. FSR 2 and 2.x are temporal: they use motion vectors, depth, and colour buffers and replace temporal anti-aliasing, with iterations 2.1 and 2.2 specifically targeting ghosting on fast-moving objects. FSR 3 adds frame generation — interpolated frames — to FSR 2-style temporal upscaling. FSR 4 is a neural-network-assisted upscaler with improved motion stability and ghosting claims, but its hardware and support scope is version- and driver-dependent and uncertain in detail: sources variously tie the full FSR 4 path to RDNA 4 and RX 9000-series hardware and cite expanding game support via Adrenalin drivers. Do not state a game count or a hardware list without a dated AMD source.

DLSS Super Resolution is temporal upscaling with a Tensor-core neural model. DLSS 3 adds Frame Generation, one synthesised intermediate frame. DLSS 4 adds Multi Frame Generation, multiple synthesised frames, and a newer transformer-based model reported to improve motion stability and thin-detail handling. Frame Generation is not upscaling: it synthesises whole frames between rendered frames. It raises the FPS counter without raising the simulation rate, and it inherits base-rate unevenness. A game may have Super Resolution on and Frame Generation off, or both on, and the two must not be conflated under a single DLSS on label.

Hardware scope, in general terms: DLSS requires NVIDIA RTX GPUs with Tensor cores. FSR 2 and 3 work across AMD, NVIDIA, and Intel GPUs, with no dedicated AI hardware required. Intel XeSS has an AI-accelerated path on Intel Arc hardware with XMX units and a DP4a fallback path for other GPUs. Game support is per game in all cases: if the game does not implement the upscaler, the driver cannot add it. Third-party DLL swapping tools exist, but they cannot add support to a game that never implemented it, so they should be mentioned only with caution, if at all. Version blindness makes comparisons meaningless: DLSS as discussed in 2021 is not DLSS in 2026, and FSR 1 spatial upscaling is not FSR 2, 3, or 4 temporal upscaling, even though they share a brand.

The Artefact Catalogue

Blur, softness, and loss of fine detail

Textures lose crispness, distant detail melts, and detail near text softens. Good implementations render the user interface at output resolution, so soft interface elements point elsewhere. The cause is simple: the internal image contains less information. At Performance or Ultra Performance, the upscaler must invent roughly 75 to 89 percent of output pixels. Low-accumulation, high-responsiveness reconstruction settings lean on the current frame and therefore look softer or smuddier on fine detail. The problem is worse when output resolution is low, or when the scene is already soft from depth of field, motion blur, or low texture settings. Upscaling cannot restore detail the renderer never produced.

Ghosting, trailing, and smearing

A faint duplicate or dark or light trail follows a moving object or character, foliage smears, or the picture takes a beat to settle after a lighting or exposure change. The cause is stale history: if motion vectors are missing, wrong, or too coarse for a fast-moving or thin object, the upscaler reprojects old colour to the wrong place and blends it in. High-accumulation settings deliberately weight history heavily to stabilise thin detail, and therefore ghost more when history is wrong. Transparency, particles, and objects that move without engine motion vectors, including some screen-space effects, are classic triggers.

Ghosting is not always the upscaler model's fault. In October 2026, a fix merged into NVIDIA's open NRI render library showed that a hard-coded pre-exposure value of 1, instead of the renderer's real pre-exposure, made FSR and DLSS history reweight wrongly when exposure changed, producing ghosting and brightness lag. The lesson is that a ghosting bug can live in game integration, not in DLSS or FSR as such, and a patch to the game or library can fix what looks like a hardware problem. That one incident should not be over-generalised into all ghosting is a bug.

Note also that the word ghosting is used for two different things, and the article must disambiguate them. Upscaler ghosting is a history trail in the rendered frame. Panel ghosting is smearing from slow pixel response, present in all content including video and desktop window drags, not only in one game with upscaling on. Readers conflate them constantly; the desktop-drag test below separates them.

Detail flicker, shimmer, and crawling

Fences, grilles, mesh, chain-link, power lines, railings, foliage, and fine patterns sparkle, crawl, or strobe as the camera moves, and distant geometry seems to boil. Thin geometry is often sub-pixel at render resolution — thinner than one render pixel. From frame to frame, a fence wire may cover a pixel, then miss it, then cover it again, so its colour oscillates. Temporal accumulation is supposed to average this out, but if the geometry moves fast, is occluded and unoccluded rapidly, or lacks clean motion vectors, the average never settles and the detail shimmers. This is the hardest case for every temporal upscaler and the reason fences and foliage recur in every comparison.

Related to this is moire on repeating patterns such as roof tiles and grilles: the render resolution cannot represent the pattern frequency, so a false pattern appears, and lower render resolutions make it worse. Mipmap bias, with negative bias values recommended by AMD in its documentation — minus 1.58 for Quality, minus 1.76 for Balanced, minus 2.0 for Performance, minus 2.58 for Ultra Performance — is one developer-side mitigation, but users cannot set it directly in most games.

Edge doubling, halos, and over-sharpening

Bright or dark outlines appear around objects, edges look crunchy, and grain looks etched. Sharpening passes — FSR's contrast-adaptive sharpening, and DLSS or FSR sharpening sliders where exposed — boost contrast at edges to counteract upscaling softness. Too much sharpening creates halos and makes noise and grain look worse. The rule of thumb from sources is to start sharpening low or moderate and increase only if the image still looks soft. More sharpening does not fix ghosting or shimmer; it makes them more visible.

Fizzle on transparencies, particles, and fast camera cuts

Smoke, rain, hair, and particle effects dissolve, flicker, or leave trails, and there is a brief wash of wrong detail for a frame or two after a camera cut or teleport. Transparencies and particles often have unreliable or absent motion vectors, and camera cuts invalidate all history at once. Good integrations flag a cut and reset history — FSR has a reset flag for exactly this — and a missed reset means one or two frames built from the previous shot's history. If the artefact lasts exactly a frame or two after a cut and then clears, that is the signature.

Settings Trade-Offs: The Decision Logic

The one rule is to choose the highest-quality mode that still reaches your frame-rate target. Do not default to Performance or Ultra Performance for maximum FPS if the target is already met: every step down buys FPS with source detail. Competitive players may rationally accept Performance artefacts for FPS; single-player players usually should not. Do not give one blanket recommendation for both.

Accumulation is a trade-off, not a free upgrade. In DLSS presets, as analysed in one secondary source, high-accumulation models, cited examples being presets C and E, blend more history, giving better fine detail and stability on thin geometry such as power lines, mesh fences, and distant foliage, but more ghosting when motion vectors are imperfect. Low-accumulation, high-responsiveness models, cited examples being presets D and F, prioritise the current frame and motion vectors, giving less ghosting but softer or smuddier fine detail. There is no single preset best for every game, resolution, and GPU, and forcing one preset everywhere, for example via driver overrides, can hurt. Preset letters and behaviour change across DLSS versions, so a preset letter must not be presented as a permanent universal recommendation, and it is uncertain whether any given reader's game exposes preset choice at all: most games do not, and it is often a driver-override or modding matter.

Sharpening is separate from the quality mode. If the image looks soft, raise quality mode first; only then touch sharpening, and back off if halos appear. Dynamic resolution, or Auto mode, adjusts render resolution live to hit a frame-rate target, and it can cause visible quality breathing, with detail level shifting as load changes. If quality seems to pulse with scene load, suspect dynamic resolution before suspecting hardware.

Frame Generation should be tested separately. If motion looks smooth but input feels laggy, or if artefacts look like interpolated smearing between frames, test with Frame Generation off and Super Resolution alone. Frame Generation also needs a decent base frame rate; a low or unstable base rate makes generated frames look worse, not better.

Finally, know what upscaling cannot fix. It cannot fix a CPU bottleneck: rendering fewer pixels does not unblock a CPU-limited game, a point one source states explicitly, so if FPS barely moves when upscaling is enabled, the limit is elsewhere, most likely the CPU. It cannot fix shader compilation stutter, VRAM exhaustion, or a broken cable or monitor. Those need their own diagnoses.

Differential Diagnosis: Upscaling Artefacts Versus the Look-Alikes

The screenshot test should be stated carefully. A screenshot or GPU capture records the rendered frame. If the artefact is in the capture, it was produced before the cable and monitor. If it is only visible on the physical monitor and absent from the capture, the fault is at or after the display output. This is diagnostic reasoning from the cable-swap, alternate-monitor, and test-pattern method in the sources, not a quoted statistic.

Do not diagnose hardware failure from artefacts alone. Upscaling artefacts, banding, and cable sparkles are all normal failure modes of healthy systems under the wrong settings. Hardware suspicion needs additional signs: artefacts in all games at native resolution, corruption in screenshots, crashes, or artefacts before the operating system loads.

Practical Checklist, in Order

1. Confirm the artefact moves with the upscaler: toggle it off to native and compare the same scene in motion, not just a still.
2. Step up one quality mode — Performance to Balanced to Quality, or to DLAA or native if FPS allows — and re-check the same scene.
3. Turn sharpening down before concluding the mode is unusable; over-sharpening mimics several artefacts.
4. Turn Frame Generation off separately from Super Resolution to isolate which feature is responsible.
5. Turn motion blur and depth of field down as a control test; they soften and smear independently of upscaling.
6. If the problem is only on one monitor, or only outside captures, swap the cable and port and test another monitor before suspecting the GPU.
7. Update the game and GPU driver: integration fixes, like the pre-exposure example, ship in patches. Remember the companion problem, though: a driver update rebuilds the shader cache, so judge image quality after the first session settles, not during the first stuttery run.
8. If one game is affected and others are clean at the same settings, treat it as game-specific first: check that title's settings, since some games expose their own upscaler version, sharpening, or anti-ghosting options, and its patch notes.
9. When to seek a check-up: artefacts at native resolution in every game, artefacts in screenshots plus crashes or corruption, or artefacts visible before the operating system loads. Those are not upscaling stories.

One last caution about promises. No FPS-gain multiplier from launch marketing should be used as a general promise: gains depend on game, resolution, GPU, and whether the game is CPU-limited. If a number is used at all, it must carry its exact game, resolution, and GPU conditions and a dated source; the safer choice is to give no multiplier. Likewise, no claim that a specific DLSS or FSR version fixes ghosting or shimmer in general should be published: improvements are per version, per game, and per scene.

For inspection and advice, the workshop is at Dali Plaza Room 1209, Shek Mun. Telephone 2186 8010, WhatsApp 6558 6806. Any repair quotation is given as a final quote after inspection. The published three-tier range is HK$400–2,800, with the final quotation subject to inspection.

Frequently Asked Questions

Why does DLSS or FSR look blurry when it is supposed to improve the picture?

Because it starts from fewer pixels than the output and must reconstruct the rest. At aggressive modes the source detail is thin, especially at 1080p output, so the reconstruction softens. Step up a quality mode first, and only then adjust sharpening. If the picture is soft at native resolution too, the cause is elsewhere, such as temporal anti-aliasing, motion blur, depth of field, or low texture settings.

What causes ghosting behind moving objects?

Stale frame history: old colour is reprojected to the wrong place when motion vectors are missing, wrong, or too coarse. High-accumulation settings ghost more in that situation because they weight history heavily. It can also be a game integration bug, as in the reported pre-exposure case, so a game-specific ghosting problem that other games do not share should be checked against patch notes before hardware is blamed.

Why do fences and foliage shimmer but walls look fine?

Thin geometry is often sub-pixel at render resolution, so a fence wire may cover a pixel in one frame and miss it in the next, and its colour oscillates. Temporal accumulation is supposed to average this out, but fast movement, rapid occlusion, or poor motion vectors prevent the average from settling. This is the hardest case for every temporal upscaler, and lower render resolutions make it worse.

Why does Performance mode look fine at 4K but bad at 1080p?

Because the starting resolution differs. At 4K output, Performance starts from 1920 by 1080, a full HD source. At 1080p output, Performance starts from 960 by 540, and Ultra Performance from 640 by 360. The mode name is the same, but the amount of source detail is not, so the advice must always be tied to output resolution.

How do I tell upscaling ghosting from panel ghosting?

Upscaler ghosting follows the 3D scene, changes with the upscaler mode, and appears in screenshots. Panel ghosting is a property of the monitor's pixel response, appears in all content including desktop window drags and video, and does not change with the upscaler setting. Drag a window quickly on the desktop: if the same smear appears, it is the panel or its overdrive setting.

Should I just use Ultra Performance for maximum FPS?

Only if you need that FPS and accept the trade. Ultra Performance at 1080p output reconstructs from 640 by 360, so artefacts should be expected. The sensible rule is the highest-quality mode that still reaches your frame-rate target, with competitive players reasonably accepting more artefacts for FPS than single-player players should.

Does upscaling fix a CPU bottleneck or shader stutter?

No. Rendering fewer pixels does not unblock a CPU-limited game, so if FPS barely moves when upscaling is enabled, the limit is elsewhere. Upscaling also does not fix shader compilation stutter, which is CPU-side pipeline building on first encounters, or VRAM exhaustion. Each of those needs its own diagnosis.