Frame generation
Also known as: multi frame generation, AI frame generation
Frame generation is a technique that uses AI and motion data to create extra frames between the ones a graphics card actually renders, boosting the displayed frame rate.
Nvidia's DLSS can now generate up to five frames for every rendered one. While the result looks much smoother, generated frames don't respond to your controls, so a game rendered at a low frame rate still feels sluggish even if the counter shows a big number.
Frame generation, explained at three levels
At home
Imagine an animator who draws the key poses of a scene, then has an assistant sketch the in-between pages so the motion flows. Frame generation works like that assistant. The graphics card draws real frames, and an AI model looks at how things moved between them and paints extra frames in the gaps. The game looks smoother, but the extra frames are educated guesses, not freshly calculated pictures.
At work
Frame generation inserts AI-created frames between those a graphics card actually renders, multiplying the displayed frame rate. Nvidia's DLSS, AMD's FSR, and Intel's XeSS all offer versions. For businesses, it lets demanding games and visualizations look smooth on high-refresh displays without more hardware. The catch is that generated frames don't respond to input, so responsiveness still depends on the real, rendered frame rate.
At home
The game renders frame one and frame two, then holds frame two back briefly. Meanwhile, an AI model studies how objects moved between them, using motion data from the game, and creates one or more in-between frames. You see frame one, the generated frames, then frame two. Holding a frame back adds a little delay, and the game only reads your controller on real frames, so it can feel less responsive than the counter suggests.
At work
Frame generation takes the two latest rendered frames plus engine data, motion vectors and depth, and runs an optical-flow or AI model to synthesize intermediate frames. DLSS 3 generates one frame per rendered frame; DLSS 4's multi frame generation generates up to three, and DLSS 4.5 up to five. Because a frame is held back, vendors pair it with latency reducers such as Nvidia Reflex or AMD Anti-Lag. A base rate near 60 FPS keeps artifacts small.
At home
Generated frames are predictions, so they fail where prediction is hard. Fast motion, objects appearing from behind others, and on-screen text such as a crosshair can smear or flicker, because the model has to invent pixels it never saw. The input delay comes from holding back a real frame to interpolate toward it. While the smoothness is real, in my opinion the honest way to compare graphics cards is rendered frames and latency, with generated frames listed separately.
At work
Integration requires accurate motion vectors and depth, and the interface rendered separately so text and HUD elements aren't interpolated. Frame pacing is critical: generated frames must be spaced evenly with real ones, which works best with variable refresh displays. Vendors ship this through SDKs such as Nvidia Streamline and AMD FidelityFX. Because every generated frame delays a real one, competitive modes should test end-to-end latency, and settings should make rendered versus displayed rates clear.
Real-world examples
At home
Aisha's new graphics card showed over 200 FPS with multi frame generation turned on. The game looked buttery smooth, but she noticed aiming felt slightly delayed compared with running it at 120 real frames.
At work
Fatima is a technical director at a Vancouver studio. Her team enables frame generation only when the base frame rate passes a set threshold, and disables it in the game's competitive mode to protect responsiveness.
How it compares
| feature | Frame generation | AI upscaling |
|---|---|---|
| What it does | Creates new frames between rendered ones | Renders at lower resolution, then sharpens |
| Effect on FPS | Multiplies displayed frames | Raises rendered frames |
| Effect on input lag | Can add slight lag | Usually reduces lag |
| Best used when | Base frame rate is already decent | Performance is GPU-limited |
Upscaling makes real frames faster. Frame generation adds frames that aren't real.
For analysts
Be cautious of performance claims built on frame generation. I'd ask for benchmarks showing rendered frame rate and latency, especially when comparing hardware for purchase.
History & origin
TVs have used motion interpolation, the 'soap opera effect,' to insert in-between frames for years, with noticeable lag and artifacts. Nvidia introduced DLSS 3 frame generation with its RTX 40 series in 2022, using dedicated hardware and game motion data for better quality. AMD followed with FSR 3 frame generation in 2023, which works on more hardware. In January 2025, Nvidia's DLSS 4 added 'multi frame generation,' producing up to three extra frames per rendered frame on RTX 50 cards, and DLSS 4.5, announced at CES 2026 and released that spring, raised the maximum to five. Intel added multi-frame generation with XeSS 3 in late 2025, while AMD's FSR still generated one extra frame per rendered frame as of mid-2026. In my view, frame generation is a great tool for high-refresh displays, and a misleading one when marketing compares generated frame rates with real ones.