mavismmg/MFGAdaUnlock-RenoDx 1.3.0

Release 1.3.0 — Adaptive Quality V3, Temporal Stability and Reflex/Pacing Lab

on GitHub

Adaptive Quality Suite V3

• Luminance + Directional V3 is now the default profile for clean installations.
• Photometric confidence now evaluates luminance and chrominance errors relative to local luminance, improving consistency across SDR, HDR, exposure changes and bright highlights.
• Motion-oriented geometry support improves handling of diagonal hair, vegetation, wires and moving silhouettes without introducing new texture reads.
• Direction-aware screen-border tapering distinguishes objects entering and leaving the image while constraining movement near corners and perpendicular edges.
• Native confidence remains the lower bound. Invalid coordinates, sentinels, NaN and infinity continue to use conservative hard rejection.
• Existing saved V1 and V2 selections remain unchanged. Components continue to fall back independently through V3, V2, V1 and the native provider path.
• Exact Adaptive Quality V3 provider validation is currently limited to DLSS-G 310.9.0 and 310.9.1. Unknown providers are not modified.

Local and Temporal silhouette stability

• Added Local Stable, using only the provider's existing 3x3 tile.
• Diagonal support, motion orientation and ambiguity decisions now use continuous transitions instead of abrupt thresholds.
• Isolated support is restricted while coherent two-sided support remains available for stable fine detail.
• Added Temporal Stable, storing only 8-bit geometry confidence—never color, depth, motion vectors or complete frames.
• Confidence loss is immediate while recovery is limited, reducing temporal grain and flicker around silhouettes.
• Symmetric interpolation phases share history buckets and exchange directions.
• History allocation is bounded to 64 MiB; 4K at 6x uses approximately 49.8 MB.
• Any API, ABI, phase, stream, context, allocation or memory-limit validation failure falls back to Local Stable.

V3.2 latency and kernel optimization

• Added separate Local and Temporal geometry cubins.
• The Local cubin contains no temporal symbols, global-memory history operations or history branches.
• The CUDA launch hook is installed only after an exact provider and Temporal cubin are validated.
• Integrations that delay-load CUDA now acquire the genuine nvcuda.dll exclusively from System32 after provider authorization.
• After the target kernel is associated, unrelated CUDA launches perform only one atomic CUfunction comparison before the original trampoline.
• Kernel names, module symbols, the 144-byte ABI and history magic are validated only during initial association.
• No hand-written SASS is used; cubins are generated with ptxas and audited with nvdisasm.

Reflex / Pacing Lab

All Reflex and pacing changes are opt-in. The default remains fully game-controlled.

• Added Game Controlled, FG-safe Off, Force On and Force On + Boost modes.
• FG-safe Off keeps Streamline's internal Reflex dependency enabled so Frame Generation remains available, while bypassing only slReflexSleep.
• Frame tokens, markers, optimization fields and other game-owned Reflex options remain intact.
• Added an experimental DXGI Waitable pacing method for games with uneven Reflex pacing.
• Waitable activation requires an existing compatible swapchain and a 120-token native safety probe.
• Timeouts, invalid handles, token errors, concurrent calls or swapchain changes immediately trigger a safe fallback.
• Added optional 0.5%–3.0% VRR/G-SYNC headroom.
• Native game limits, Output FPS Cap, Latency Guard and VRR headroom always use the most restrictive applicable limit.
• The native slReflexSleep wrapper retains a lock-free direct-call fast path when the laboratory is inactive.

Interface and diagnostics

• Added requested/effective Reflex mode, pacing state, waitable probe, refresh, VRR and cadence diagnostics.
• Added CUDA hook, target association, temporal phase, confidence-history and exact fallback diagnostics.
• Fixed clipped combo labels, help markers and long provider/status text at narrower overlay widths.
• Long diagnostic values now wrap within their table columns.
• Application cadence measurements are explicitly separated from externally measured displayed/generated frames.

Compatibility notes

• Reflex/Pacing Lab and DXGI Waitable remain experimental and should be validated per game.
• Use Game Controlled + Native Reflex Sleep for maximum compatibility.
• Use Local Stable if a game or driver cannot validate the Temporal CUDA path.
• Dynamic MFG continues to require the validated DLSS-G 310.9.1 + Streamline 2.14.1 D3D12 stack.
• No automatic game-specific pacing override is applied by executable name.

Validation

• All 20 automated tests passed.
• Clean Release build completed successfully.
• Local and Temporal geometry cubins use 40 registers, 7,776 bytes of shared memory and zero stack, local memory or spills.
• The Local cubin contains no global-memory history loads or stores.
• The Temporal path contains exactly two confidence-byte reads and two confidence-byte writes per pixel.
• The V3 warp program uses 48 registers with zero spills.
• Native slReflexSleep wrapper remained within the latency microbenchmark limits.
• No NVIDIA DLL or complete provider package is redistributed.

Credits

• Original ReShade/RenoDX addon: Dreamt
• Original RTX40 MFG unlock and temporal-correction research: dashdogy
• Blackwell-kernel research and rebuild workflow: Matias Lombo
• Public qualityValidWarp research that informed the independently implemented Validated Warp Blend path: Tony Joaca
• Monster Hunter Wilds first-launch pacing investigation and testing: Darkalibur
• Star Wars Outlaws compatibility testing: harddaysmike
• Compatibility, quality, UI, diagnostics, validation and maintenance work in this fork: mavismmg

This remains a fork of Dreamt's original project. Existing repository credits and the original license are preserved.

SHA-256 (renodx-mfgunlock.addon64): 7D742480613C58DC0EB698BF5DC76F2B32A6B4C607B2FACE5B87940494B9B229