Video summary
DLSS 5 is NOT What We Expected
Main summary
Key takeaways
Tech/Product concepts discussed: DLSS 5 “unsanctioned” demos and how it changes image rendering
DLSS 5 positioning & controversy
- Nvidia markets DLSS 5 as “AI superpowered” graphics, but the hosts frame it as controversial—potentially acting more like an AI filter than traditional DLSS’s goal of improving performance.
- Even if Nvidia claims it preserves artistic intent, the implementation is still questioned.
- Performance cost: the hosts report roughly a ~50–60% performance hit on their system, with even a high-end RTX 5090-class setup struggling.
How DLSS 5 works (pipeline)
DLSS 5 is designed to fit real-time timing by:
- Using a rendered image plus color and motion vectors at the end of the pipeline
- Applying an AI uplift after those inputs are available
Control scheme exposed in modded demos
The discussion centers on modders enabling DLSS 5 in games that provide color and motion vectors (an example mentioned is a leaked/embargo demo for NBA 2K27 repurposed via modding).
In the modded version, reviewers appear to access developer controls that regular users may not:
- Users get a simple on/off
- Developers can adjust:
- 3 DLSS 5 models (selectable per scene): Default, Natural, Cinematic
- Intensity (described as largely “turning it”)
- Tone (color/lighting tint)
- Structure (interaction details, including subsurface scattering)
Key evaluation approach
They rely on pixel-peeping comparisons (toggle on/off) and focus on:
- Whether lighting matches intent (or becomes “more natural but not necessarily better”)
- Face artifacts (sometimes uncanny results)
- Changes to NPCs and character detail fidelity
- FPS impact and stability, including hitching in at least one title
Game-by-game findings (image quality + issues)
Cyberpunk (4K baseline / DLSS “Balanced” example)
Positives
- Lighting can look more natural (e.g., changes to shadow/ambient behavior).
- Subjective improvements include:
- More realistic lighting falloff
- Better perceived detail in clothing/geometry
Negatives / concerns
- They call out bad/incorrect results such as wibbly/wobbly lighting around the neck and other RT/fake lighting artifacts.
- DLSS 5 can materially alter character appearance:
- Different models (Default/Natural/Cinematic) change ambient occlusion, face lighting, and even perceived anatomy/detailing.
Broader character quality (especially NPCs)
- One reviewer claims DLSS 5 delivers the best NPC quality they’ve ever seen—random NPCs can look closer to main characters.
- Nvidia’s stated intent is developer-level control (different filtering amounts for different objects), which the hosts see as a key feature—DLSS 5 could potentially separate objects and treat them differently.
Final Fantasy VII Rebirth
Mixed results
- Some improvements resemble texture-like gains (e.g., jacket stitching, armor scratches, background detail).
Major critique
- Faces look wrong: attempts to make anime characters feel real can produce Instagram-face / chiseled cheeks unnaturalness.
- They liken it to the “realistic Mario” meme—defined item detail vs “blob” detail—but note it can become uncanny.
“Claire Obscure” (another tested game)
- Improvements appear in some clothing detail, but overall visuals may look:
- Flattened, or
- With overpowering shadows
- They feel the “dreamlike” artistic look can be lost when DLSS 5 shifts toward more “natural” colors.
- Reported hitching/stutter occurs when toggling DLSS 5 tone in this game.
Close-up facial detail / pores / lip gloss look
- In close-up comparisons, DLSS 5 can make skin look extremely real (e.g., pores and lip gloss/chapped lips impressions).
- But it can also become excessive:
- Faces may appear to shift in age (comments like “kid looks 47/80”)
- Freckles/skin microdetail can read as scary/uncanny, depending on the character and implementation
007 / “This looks so good by default” then DLSS uplift
- They show a strong uplift when enabled (jokingly compared to a “Steam window”).
- Still emphasized as early/unsanctioned and possibly not representative of final release behavior:
- The demo behavior is taken from another game/mod context.
Overall conclusions / “review verdict” style takeaways
Best framing given by hosts
- They describe this as the worst version DLSS 5 will ever look—their testing is very early and imperfect.
- They estimate mainstream consumers may not see meaningful results on typical hardware for about ~3–5 years (with a joke suggesting “3–5 days”).
- They expect developers/modders will have time to refine it, and note Nvidia seems focused on current-gen GPUs (with hints about 4000-series mod support).
Likelihood of adoption
- For their test set:
- Only a minority of cases looked “incredible”
- Some were “okay”
- Some should not be used
- They state they would not buy a GPU solely to enable DLSS 5 due to the performance hit and inconsistent outcomes.
Main speakers/sources
- LTT / Linus Tech Tips-style hosts:
- Elijah (explicitly named)
- David (explicitly named)
- Plof (mentioned in the RTX Remix follow-up segment)
- Source being discussed: Nvidia (DLSS 5 claims/positioning)
- Modders: referenced as enabling DLSS 5 in other games and exposing “developer controls” via mods