Can You Run Deep-Live-Cam Without a GPU?

Every guide we found says CPU mode works but is slow. None of them says how slow, so we put a number on it.

Does Deep-Live-Cam work without a graphics card?

Yes, it starts and swaps faces on the processor alone. Its README says the manual install "will be slower as it utilizes the CPU", and that without a GPU you simply run python run.py. The GPU steps (CUDA for Nvidia, CoreML for Apple silicon, DirectML for Windows, OpenVINO for Intel) are all optional extras on top.

What it needs either way: Python 3.11 to 3.14 (3.14 recommended), git, ffmpeg, the Visual Studio 2022 runtimes on Windows, and two model files in the models folder: inswapper_128_fp16.onnx (the swap model) and gfpgan-1024.onnx (the optional face enhancer). Guides that still mention GFPGANv1.4.pth or CUDA 11.8 are out of date; the README now links the ONNX enhancer and CUDA 12.8.

How fast is Deep-Live-Cam on a CPU?

About one frame a second. On a 4-core Intel Xeon cloud machine (2.8 GHz, no graphics card), the inswapper_128 pipeline took a median of 1,192 ms per frame (0.84 frames a second), and 1 frame in 20 took 1,426 ms or more. A smooth 30 fps picture needs 33 ms. Our GPU server swaps a live frame in 24 ms.

Where the time goes: finding the face took 130 ms, pasting it back took about 10 ms, and the swap model itself took 1,043 ms. That is why the usual CPU tips barely help. Halving the detector size cut detection from 130 to 59 ms, which saved 7% of the frame. Nothing you set in the app makes the swap model smaller.

What does a CPU swap look like?

The same as a GPU swap, just much later. The processor runs the same model and gets the same picture; speed is the only difference.

The swap kept the camera's smile, teeth, hairline and sweater and took the eyebrows, eyes and beard from Ezra. In a live call, though, you'd see this picture change about once a second, with your lips a full second behind your voice.

Do more CPU cores help?

Yes, up to a point. With the swap model alone, one thread took 3,282 ms, two took 1,831 ms and four took 1,083 ms. Each doubling saved less than half. Deep-Live-Cam picks the thread count for you on a CPU: your core count minus 2, at least 4 and at most 16 (--execution-threads overrides it).

A recent 8- or 12-core laptop will beat our 4-core cloud machine. But to reach 30 frames a second it would have to be about 36 times faster, and even a choppy 10 frames a second would take 12 times.

Which settings make Deep-Live-Cam faster on a CPU?

Three things helped in our test, and one didn't. These follow Deep-Live-Cam's README and code; we timed the models, not its installer.

1. 1

   Leave the face enhancer off

   It's the biggest cost on a CPU. GFPGAN 1.4 added 1,854 ms per frame (0.33 frames a second in total). If you want some sharpening, GPEN-256 cost 625 ms, under half as much. The README's gfpgan-1024.onnx works on a larger face than the 512 px GFPGAN 1.4 we timed, so expect it to cost more, not less.

2. 2

   Pick the smaller face detector

   Start with python run.py --execution-provider cpu --det-size 320. It saved 87 ms a frame for us. Try 160 only if your face fills the frame: a smaller detector can miss small or distant faces.

3. 3

   Give it every core

   Close other heavy apps, and if you want to test more threads than the default, add --execution-threads with your core count.

4. 4

   Don't bother swapping fp16 for fp32

   On a CPU the two model files ran at the same speed (1,174 ms vs 1,192 ms). Without an Nvidia card, Deep-Live-Cam loads inswapper_128.onnx (fp32) when it's in the folder and falls back to the fp16 file the README links, so keep the fp16 one: it's half the download (278 MB vs 555 MB).

What is a CPU good for, then?

Video files, where nobody waits for the next frame. Deep-Live-Cam's command line takes a face, a video and an output (-s face.jpg -t clip.mp4 -o out.mp4). At our speed, one minute of 30 fps video is 1,800 frames, about 36 minutes of work without the enhancer and about 91 minutes with GFPGAN. A 10-second clip takes about 6 minutes. In our runs, loading the models took another 6 to 25 seconds before the first frame.

If you'd rather not wait, Video Face Swap is a browser tool that swaps the face in every frame of a short video for the face you upload, keeps the sound, and gives you an MP4 back at the original size. It runs on our GPUs, so your computer doesn't matter.

How do you face swap live without a GPU?

Run the swap on someone else's GPU. You can rent a cloud GPU and run Deep-Live-Cam there, but then you have to get the picture back into your call yourself. Or use a tool built that way from the start.

Live Face Swap is a real-time face swap that lets you look like anyone you choose: try it with your webcam on facecam.ai, then use it as a virtual camera in Google Meet, Zoom on the web, Discord in the browser and other sites in Chrome with the extension. Your laptop only sends camera frames and shows the result; in our benchmark a swapped frame reached the screen in about a fifth of a second. You pick Facecam Virtual Camera as the camera in the call, with no Python, models or OBS. It's on every paid plan, from $10 a month, at 4.6 credits a minute. How real-time face swap works explains the round trip.

Where Deep-Live-Cam still wins: it's free, works offline once installed, and its README points to OBS to get the picture into a stream. Facecam needs an internet connection, and runs in Chrome or, for desktop apps, in Facecam for PC and Mac. Our Deep-Live-Cam alternative page compares the two in full.

What didn't we test?

We didn't install or run the Deep-Live-Cam app. We ran the models it uses through our own pipeline, so its live loop, which has its own shortcuts and draws a preview window, may be a little faster or slower than our numbers. Our machine was a 4-core cloud server; we didn't test a desktop processor, Apple silicon or an Intel chip with OpenVINO, which the README supports. We timed GFPGAN 1.4 at 512 px, not the 1,024 px file the README links. And we didn't test the pre-built CPU version offered on Deep-Live-Cam's site.