Pavise Game Optimizer Explained: 96 vs 59 FPS, Zero Injection [2026]
Friday, September 11, 2026Updated September 11, 2026: Pavise by bdth is a Chinese C# game optimizer that hands CPU, I/O, and scheduling resources back to your game: process priorities, core affinity, GPU scheduling, input latency fixes, and NVIDIA tuning, all local, all reversible, never injecting into game processes. It holds 448 stars and 16 forks across 12 commits, publishes its own benchmark rig results including 33-to-95 FPS background improvements, and its license bans selling it after scam copies appeared. Documentation spans Chinese, English, and Japanese. Here is every mode, the measured numbers, and the safety boundaries.
Table of Contents
- What Pavise Is
- The Benchmarks: 59 to 99
- The 4 Modes
- CPU, Cores, and the 33-to-95 Fix
- GPU and NVIDIA Tuning
- Input Latency Work
- Memory Cleanup: Measured and Discouraged
- Safety Boundaries
- License and Scam Warning
- Bottom Line
- Frequently Asked Questions
1. What Pavise Is
Pavise by developer bdth is a Windows gaming performance tool that returns CPU, I/O, and scheduling resources taken by background processes to the game. Built in C# with WinForms, documented in Simplified Chinese with full English and Japanese translations, the repository holds 448 stars, 16 forks, 1 watcher, and a single issue across 12 commits. Its defining architectural promise: runs locally, installs no service, uploads no data, never injects into game processes, never touches game memory or files, and every change is reversible.
The no-injection design is the whole security story. ReShade add-ons, DLL proxies, and driver tweaks all touch the game or its rendering path, which is why anti-cheat systems flag them. Pavise operates one level out, managing priorities, affinities, services, and OS settings around the game while the game process itself stays pristine. Anti-cheat has nothing to detect because nothing enters its jurisdiction.
Setup follows the same restraint. Builds compile with Windows' built-in .NET Framework compiler, no Visual Studio, no package restore. The executable runs from the tray, requests admin rights only because adjusting other processes requires them, and stores data in AppData with switches in the registry, or portable beside the binary with a Pavise.portable marker file. Launching adds one version check against GitHub and uploads nothing.
2. The Benchmarks: 59 to 99
The README publishes rig results instead of promises. Test conditions: an i7-9750H laptop running WeChat, Clash, QQ, and QQ Music in the background, scored against a clean system baseline of 100. Averaged across three scenarios, no Pavise scored 59, Normal mode 60, Competitive 96, and Extreme 99. The author states the mechanism plainly: Pavise creates no performance, it returns the background-stolen share to the game.
Two details elevate these numbers above typical optimizer marketing. First, the author publishes the case where stronger modes lose: Extreme scores only 71 in high-load scenarios against Competitive's 93, with the explicit conclusion that suppression strength and performance are not positively correlated and the default preset is the gentlest tier. A developer showing where the max setting fails is reporting, not selling.
Second, the gains concentrate where contention exists. Background-heavy machines with visible CPU or disk fighting improve markedly; clean systems or fully GPU-bound games improve little. The 1% low frame improvement of 77 to 96 percent under full CPU load, attributed to key-thread identification, targets exactly the stutter that averages hide. Users are told to compare with the toggle on identical games and scenes rather than trust any number.
3. The 4 Modes
Normal mode applies light suppression to windowless background processes plus extra pressure on persistent resource hogs, leaving everything in use untouched. Competitive suppresses everything except the game, including windowed apps, keeping only the current foreground program and its children. Extreme adds service-host processes to background cores on top of Competitive. Custom exposes background suppression, network, notifications, and items individually.
The mode ladder encodes a philosophy of escalating consent. Each step suppresses more of the system and says so, with Competitive's foreground-only rule the clearest expression: if you are looking at it, it lives, everything else yields. Custom exists for users who read the feature list and want surgical control rather than presets.
Game detection runs automatically during play, survives alt-tab and minimization, and handles launcher chains: League of Legends-style launches get confirmed once, then the real game binary is recognized directly. Launchers, updaters, crash reporters, and anti-cheat processes never count as games. Exiting restores every change from the session record, and a crash mid-session resumes restoration on next launch.
4. CPU, Cores, and the 33-to-95 Fix
Game processes receive high priority, elevated disk I/O and GPU scheduling priority, and dedicated cores, while background processes downgrade or migrate per mode. CPU partitioning adapts to hybrid architectures, X3D layouts, and multiprocessor groups, skipping core splitting entirely on 6 cores and under. Manual per-core assignment offers all-core, no-hyperthreading, performance-cores-only, and invert presets, applied only on confirmation, though the author notes most users should leave the scheduler alone since game threads land on P-cores by default.
The standout feature is background binding contraction: already-suppressed background processes shrink from full-machine spread to a few physical cores, attacking concurrent memory-bandwidth contention that priority and efficiency modes cannot reach. Placement follows the user's game cores in real time and dodges the game's L3 entirely on multi-CCD chips. Rig result: 33 to 95 FPS in the measured case. That single fix addresses a bottleneck most optimizers never name.
Service yielding moves designated service processes to background cores during matches without stopping anything, routing rather than killing. Kernel-blocked games get fallback elevation applied by the system at process creation, effective next launch. Every mechanism prefers moving work over stopping it, which is why restoration after exit is complete rather than approximate.
5. GPU and NVIDIA Tuning
Background GPU yielding lowers scheduling priority for background processes using the card. NVIDIA deep tuning sets maximum-performance power, DLSS 4 Transformer overrides, per-game ReBAR, Ansel injection removal, and battery frame-rate cap removal, every original snapshotted and restored on toggle-off. On dual-GPU machines, suppressed background programs using the discrete card get moved to integrated graphics.
The snapshot discipline matters. Registry and driver tweaks that tools apply and forget are how systems accumulate mystery states across years of optimizer experiments. Pavise reads back each write where possible, counts only verified writes as success, and restores originals when features switch off. Upgrades auto-restore values written by removed features, cleaning up after the project's own past experiments.
Removed features get their own documentation section, a rare honesty. MSI mode, low-latency mode, background hard frame caps, windowed-game optimization, file preheating, and CPU 0/1 masking were all implemented and removed for measured ineffectiveness, device-bricking risk, or semantic error, including the discovery that driver background frame caps actually throttle the game itself when focus leaves it. Failed experiments stay documented so nobody re-adds them.
6. Input Latency Work
Filter, sticky, and toggle keys get disabled because Microsoft defines them as ignoring brief keypresses, which is latency by specification. System suspension of mouse and keyboard devices is blocked to cure the first-input drift after idle, touching only input devices and leaving USB drives and audio alone. Modified input queue lengths get repaired, pointer precision enhancement, the OS mouse acceleration curve, switches off.
High polling-rate mice get interrupt isolation: 4K and 8K mouse interrupt storms route away from game cores, with an upfront cost warning and the note that 1000Hz and below gains little. Storage controller read-completion interrupts move to idle cores. These are DPC and ISR level adjustments, the layer where microseconds actually live, and the system-health page names the top offender per machine using kernel ETW tracing.
The project declines most registry folklore. Widely circulated mouse registry tweaks act at fractions of a millisecond while render queues dominate latency, so Pavise reports items like Bluetooth mice and 125Hz rates without modifying them. The health page flags, the user decides. An optimizer that refuses to optimize on request earns more trust than one that tweaks everything.
7. Memory Cleanup: Measured and Discouraged
All three memory features ship off, and the README recommends keeping them off for machines with ample RAM. The justification is a measurement table on a 64GB rig that more optimizer authors should be forced to read. Emptying all process working sets took 1894.9ms for 5191MB freed. Clearing the full standby list took 1380.6ms, destroyed 18894MB of system cache, and gained only 382MB of usable memory. Flushing modified page lists took 8 seconds. Merging physical pages took 6.2 seconds.
The conclusion is devastating to the genre: emptying standby converts cached-available memory into empty-available memory, barely changing the total while forcing subsequent game loads back to disk. The three features exist only for memory-starved low-end machines. Pre-match low-priority standby trim costs 13ms and skips valuable caches. Post-load background working-set recovery at 30 seconds returns pages to standby per process without global commands. Threshold-triggered standby clearing, an ISLC equivalent, fires only above 1GB standby with under 1GB available, targeting lock contention micro-stutter rather than capacity.
The section cites Mark Russinovich's 2004 Memory-Optimization Hoax essay as prior art. A Chinese gaming tool grounding its memory chapter in a 22-year-old Microsoft essay about why RAM optimizers are scams is the strongest credibility signal in the repository. The feature exists, the author measured it, the author tells you not to use it.
8. Safety Boundaries
| Boundary | Rule |
|---|---|
| Anti-cheat and match hosts | Never suppressed, in any mode, no override exists |
| Windows core services | Never suppressed, boundary immune to all switches |
| Game accelerators, other users | Excluded from suppression entirely |
| Game processes | Never injected, memory and files untouched, verified writes only |
| Restoration | All changes reverted on exit, crash-safe resume, one-click reset page |
| Network | One GitHub version check at launch, zero uploads, fully local |
9. License and Scam Warning
| Term | Detail |
|---|---|
| License | Pavise license: open source, free use, modification, and gratis distribution |
| Sale ban | No money from distribution in any form: copies, keys, paywalls, subscriptions, tip gates |
| Distribution duty | Keep license and author credits, mark modifications, inform recipients of the ban |
| Scam warning | Paid copies are fraud, latest builds free in QQ groups and WeChat, author lists contacts |
10. Bottom Line
Pavise is the rare optimizer that publishes losing benchmarks, documents removed features, measures its own memory tools into irrelevance, and bans its own sale after scammers moved in. The no-injection architecture makes it anti-cheat safe by construction. The restoration guarantees make it risk-free to try. Competitive mode's 96 against a 59 baseline, on a cluttered laptop, is the number that matters, with the author's own warning attached: clean systems and GPU-bound games gain little. Cluttered systems gain afternoons. Free forever, in three languages, from the groups, never from a seller.
Frequently Asked Questions
Q: What does Pavise do?
A: It returns CPU, I/O, and scheduling resources from background processes to the game through priorities, core affinity, GPU scheduling, input fixes, and NVIDIA tuning. Local only, no injection, no services, everything reversible.
Q: Is it safe with anti-cheat?
A: By construction. Pavise never injects into game processes or touches game memory and files, and anti-cheat processes sit inside the never-suppress boundary. Standard optimizer risk does not apply to its architecture.
Q: What performance gains are measured?
A: On an i7-9750H laptop with chat and music apps running: 59 baseline, 60 Normal, 96 Competitive, 99 Extreme averaged, with 1% lows up 77 to 96 percent and a 33-to-95 FPS background-contraction case. GPU-bound and clean systems gain little.
Q: Should I enable the memory cleanup features?
A: No, unless RAM is genuinely tight. The author's own measurements show standby clearing destroys 18GB of cache for 382MB gain with multi-second stalls. All three ship off and should stay off on ample-memory machines.
Q: Which mode should I use?
A: Competitive for most gamers, since it preserves only the foreground program. Extreme adds service routing with weaker high-load results. Normal barely moves the needle. Custom serves users who read the feature list.
Q: Where do I download it safely?
A: GitHub releases and the author's free QQ groups and WeChat. The license bans all paid distribution, and the author warns paid copies are fraud demanding refunds. Never buy Pavise from anyone.
Byline: Indie Kings | September 11, 2026
Labels: PC Gaming
Related: Windows 11 October Update Auto-Enables Memory Integrity | Windows 11 8GB Diet Promise vs. Reality | Steam Deck Hits 30,000 Verified Games