Skip to content

PushFPS v3.2.10

Choose a tag to compare

@github-actions github-actions released this 20 Nov 08:01

What's new in 3.2.10: the last three PushFlow 2 switches

With this version every part of the PushFlow 2 plan has its switch. All of them start off, and with
all of them off PushFPS makes exactly the frames it made before (the tests compare them).

You find them in Professional → Frame generation, under the PushFlow tier. This version adds three.

Dual: true frames for recordings and streams. With Predict on (3.2.9), your screen shows forecasts, and
recordings, instant replay and streams recorded what was on screen, forecasts included. With Dual they get
the true picture instead: the real frames, and the true in-between frame of each pair. Predict has already
worked that frame out, so it costs almost nothing.

  • What you see, and its delay, do not change.
  • It acts only while Predict runs, at 2X. At 3X and 4X recordings stay as played.
  • The recording runs one frame behind the screen, and its sound is one frame ahead of its picture (about
    33 ms at 30 FPS).
  • Comparison clips still record what is on screen.
  • No automatic test here makes an actual recording with it. The frames handed to the recorder are
    checked; the file is not. Please look at one recording before you rely on it.

Response: a measurement first. The idea of Response is to turn the picture a little from the right
stick before the console's own frame arrives. Nobody has shown that this can be done from a video signal,
so the first step only measures, and changes nothing on the picture: does the picture's own motion follow
the right stick, how strongly, and how late?

  • It works only where the controller goes through PushFPS: PS5 Remote Play. A controller plugged into
    the console cannot be seen, and the row says so.
  • It looks at the right stick's two axes and at the picture's own motion (from PushFlow, while Predict
    runs). No button is read, nothing is written to disk, nothing is sent anywhere, and everything is
    cleared when you switch it off or change the source.
  • Play for a few minutes and turn the camera a lot. The row then says whether the picture follows the
    stick, by how much and how late. That answer, from your game, is what decides whether Response is worth
    building further.
  • Use the Quality tier for it if your card allows: in our test the Fast and Balanced tiers' model
    found a panning fine texture moving at only a fifth of its real speed, so "how much" is less reliable
    with them. "Whether" and "how late" depend on the tier far less.

Memory: a measurement first. The idea of Memory is to bring back detail that earlier real frames
showed. That only helps where neighbouring real frames really differ, so the first step measures it: where
the motion is known, how far apart are two real frames once they are lined up? A capture card's picture
is not compressed and should hardly differ; a streamed picture should differ more. The row shows the
figure and what it means. Nothing on the picture changes.

Both measurements need Predict running, and both are in the performance report (Settings → Performance
report
), so a report from your setup tells us what no test here can.

If something goes wrong, PushFPS.exe --safe-mode starts with every experiment off, for that run.

What was tested, and how

  • Built and tested automatically on Windows (GitHub's build machines, software renderer). Dual: through
    the whole frame generation pipeline (the recorder's frames are the real frame before and exactly the true
    in-between frame, for every pair) and through the whole engine with a live source (they are handed over
    at the right moments, and the screen is as with Predict alone). Response: its arithmetic on made-up
    sticks and pictures with a known answer, and the motion it reads from a panning and a still picture.
    Memory: the same picture twice, and a picture made slightly brighter.
  • Not tested on a real graphics card, capture card, console or controller, and no recording was made
    with Dual. Response and Memory have never seen real input: finding out what they say on real setups is
    what this release is for.
  • Nothing in capture, recording, streaming or the other frame generation methods changes for people who
    leave the switches off.

Windows Defender note. Defender may flag new PushFPS downloads as Trojan:Win32/Tecabans.STV!cl. This is a false positive from its cloud check on new, unsigned files, and it has been reported to Microsoft. Always check the zip against SHA256SUMS.txt on this page.

Credits. PushFPS is built on Veyra by Likely7. PushFlow runs RIFE by hzwer. ANR frame generation is by Apex Compute.

Install

  1. Download PushFPS-3.2.10-win64.zip below and extract it.
  2. Double-click Install-PushFPS.cmd.

The installer downloads Veyra's official runtime package (about 750 MB, NVIDIA or AMD to match your
graphics card) from Veyra's GitHub release, checks its SHA-256, and installs PushFPS with Start menu
and desktop shortcuts. No administrator rights needed.

Built and tested automatically on Windows. Please report anything that doesn't work in Issues.

SHA-256: 2792F6BE5C9E3E7EDDF0AE7E8E34AC840CCE58B16A204C06AFAB91D02D6DD4EE PushFPS-3.2.10-win64.zip

Setup guide

PushFPS-Setup-Guide.txt has the recommended settings to start from, how to tune them, the
settings for PS5, Xbox and Switch, the Windows settings and when to use HDR. Also on the website:
https://dosukasol.github.io/PushFPS/#guide

OBS Studio plugin

PushFPS-OBS-Plugin-1.2.0-win64.zip adds a PushFPS source to OBS Studio 28+ (Windows): your PushFPS picture,
frame generation included, in any scene. Extract it and double-click Install-PushFPS-OBS.cmd.

SHA-256: b25a99599bba02d7dce41c3e257b404b54d56975bcf3f5296de87e297e2e4b70 PushFPS-OBS-Plugin-1.2.0-win64.zip

Linux and macOS (preview)

PushFPS also runs on Linux and macOS with its own frame generation, Frame Pacing Repair and Live HUD
Lock (no vendor components). The packages (linux-x86_64 folder or AppImage, macos-universal app)
are attached to this release when their build finishes. See the README for what is tested.

Checking the download

SHA256SUMS.txt lists the SHA-256 of every file; *-components.txt lists the pinned components
this build was made from.

License and source

PushFPS is free software under the GNU GPL-3.0 and comes with no warranty. It is a modified version of
Veyra by Likely7 (see NOTICE.md). Corresponding source:
the v3.2.10 tag of this repository (also attached by GitHub as Source code); the build recipe is
.github/workflows/build-windows.yml at that tag. The streaming libraries built into PushFPS
(chiaki-ng, AGPL-3.0 with the OpenSSL exception; moonlight-common-c, GPL-3.0; and the vcpkg recipes of
OpenSSL, Opus, SDL3, libdatachannel and the other static libraries) are in the attached
streaming-source archive.