As rendering lead for a game shipping on several hardware tiers, how do you set the anti-aliasing and upscaling policy, including each tier's internal resolution?
answer
- frame time and the look first
- one temporal pipeline, several back ends
- internal resolution per tier
- motion vectors as a content rule
basics
~20 sBuild one temporal pipeline every tier shares, set each tier's output and internal resolution from its GPU time budget, put rival upscalers behind one interface, keep a non-temporal fallback, and make motion vectors and masks a content rule.
solid answer
~50 sI start from each tier's frame-time target and the game's look: thin geometry, foliage and specular detail decide how much anti-aliasing has to do. For most projects the answer is **one temporal pipeline** shared by every tier: jitter, motion vectors, depth and reactive masks feed either native TAA or a temporal upscaler, with DLSS, FSR and XeSS behind **one internal interface** because they consume the same inputs. Each tier gets a fixed **output resolution** and an **internal resolution** or a dynamic range, chosen by profiling the hardest scenes, with a floor below which reconstruction visibly fails. I keep a non-temporal **fallback** where players need it, such as SMAA for motion-clarity purists, and I fix the **content contract**: everything that moves writes motion vectors, and QA captures every mode on the scenes that break temporal methods.
go deeper
Know the families, MSAA, post-process AA, TAA and temporal upscalers, and that rendering below the output resolution saves shading time.
Explain what each method costs and needs, especially the inputs a temporal pipeline requires, and how a per-axis scale relates to the pixel count shaded.
Profile each tier, find where reconstruction fails on the hardest scenes, and turn that into resolution floors and per-mode regression captures.
Own the trade between clarity, stability, frame time and test cost across tiers, decide what is locked or exposed to players, and make the inputs a content standard.
## What the policy decides An anti-aliasing and upscaling policy answers a handful of questions once for the whole project, so that every team builds against the same assumptions: - Which **methods** exist in the renderer, and which one is the default on each hardware tier. - Each tier's **output resolution** and **internal resolution**, fixed or dynamic, with its floor. - What is **locked** on fixed hardware and what is **exposed** to players on open platforms. - What **content** must provide so the chosen methods work: motion vectors, masks, minimum feature sizes. - How the result is **tested**, and on which scenes. There is no single right answer. The same game might reasonably ship native TAA everywhere, or upscaling everywhere, depending on its look and its frame-time targets. ## The forces on the decision - **Frame time**: the share of each tier's budget the frame can spend on shading pixels, and how much an internal resolution below output buys back. Pixel count scales with the square of the per-axis scale. - **Content**: thin railings, wires, dense foliage and sharp specular are what crawl; a stylised game with flat shading needs far less help. - **Motion clarity**: temporal methods trade softness and occasional ghosting for stability. Fast competitive play and some players reject that trade. - **Memory and latency**: history buffers at output resolution, extra inputs such as motion and masks, and the frames history needs to converge. - **Shading architecture**: renderers that keep many per-pixel targets make MSAA expensive, which narrows the menu; how each architecture copes is a separate topic. - **Availability**: the rival upscalers do not all run on every GPU, and each needs integration upkeep. - **Test cost**: every mode multiplies the QA matrix, so each one must earn its place. ## The options compared | Method | Shading cost | Stability in motion | Sharpness | What it needs | |---|---|---|---|---| | MSAA | Edges only, plus per-sample memory | Good on edges, none for shading | High | A design that tolerates multisampled targets | | SMAA or FXAA | Very low | Poor | Medium | Nothing beyond the final image | | Native TAA | Low | Good | Softer | Jitter, depth, motion vectors, masks | | Temporal upscaler | Below native, sometimes far below | Good | Close to native when inputs are right | The same inputs, plus mip bias and output-resolution history | | Supersampling | Several times native | Good | Highest | Spare GPU time | ## A worked tier policy The figures below are this scenario's, not a recommendation for any hardware. | Tier | Output | Internal | Method | |---|---|---|---| | High | 3840x2160 | 2560x1440 (about 44% of pixels), dynamic down to 1920x1080 (25%) | Temporal upscaler | | Mid | 2560x1440 | 1707x960 (about 44%), dynamic down to 1280x720 (25%) | Temporal upscaler | | Low | 1920x1080 | 1280x720 (about 44%), floor 1280x720 | Temporal upscaler; SMAA fallback | | Open platforms | Player's choice | Resolution-scale slider | Native TAA, any available upscaler, SMAA | How the table was reached: 1. Profile the hardest scenes on each tier at native output to find the shading cost to recover. 2. Lower the internal resolution until the frame fits, then go lower until reconstruction visibly fails on railings, wires and particles over a bright sky. The floor sits above that point. 3. If a tier still misses its target at the floor, cut content cost on that tier instead of resolution. 4. Lock the defaults on fixed hardware; expose methods, the upscaler choice and a sharpening control on open platforms. ## Rules that make the policy hold - **One upscaler interface**: DLSS, FSR and XeSS sit behind a single internal API fed by the same jitter, depth, motion vectors, exposure and masks, so adding or dropping one is integration work, not a renderer change. - **Motion vectors are mandatory** for everything that moves; missing ones are bugs, not polish. - **Reactive masks** are authored by the effects team for particles, glass and animated materials. - **Mip bias is derived** from the internal-to-output ratio automatically, never tuned per material. - **The HUD composites at output resolution**, after the upscale. - **Regression captures** of railings, wires, foliage in wind, particles over sky and camera cuts are recorded per mode and per tier and reviewed when the renderer or an implementation changes. ## When to revisit New tiers, a new upscaler release, content that changes the risk (a forest level full of thin branches), or player reports of blur and ghosting each reopen the policy. Each change is weighed against the test matrix it adds.
- Why keep a non-temporal option at all if TAA and upscaling look best in screenshots?Temporal methods trade motion clarity for stability. Some players, and fast competitive modes, prefer a sharper image with some crawling edges, and some content, such as fast particles over busy backgrounds, still ghosts. A cheap SMAA mode costs little to keep, serves those players and gives QA a reference image without history to compare against.
- How do you choose the floor for a tier's dynamic internal resolution?Render the worst content at decreasing scales and find where reconstruction visibly fails: thin lines break up, in-world text becomes unreadable, particles smear. Set the floor above that point. If the frame still misses its target at the floor, reduce content cost on that tier rather than resolution. The floor differs per tier because the output resolution differs.
- Integrate one upscaler implementation directly; the others can be swapped in later for free.
- Use the same internal-resolution percentage on every tier and in every scene.
- Anti-aliasing is a settings-menu toggle, so content teams need not care about it.
- MSAA is always the right answer because it gives the sharpest image.
- A sharpening pass can recover any internal resolution, however low.