Important
After installing or updating the addon, allow first-launch DLSS-G kernel compilation to finish, then fully restart the game before evaluating image quality or performance.
Adaptive Quality V3.4 — Temporal Inpaint Stability
- Added V2 Compatibility, Local V3 and Temporal V3 inpaint modes.
- Temporal V3 is now the default when the setting is absent. Existing saved choices are preserved; invalid values fall back to V2 Compatibility. Changing modes requires a game restart.
- Temporal stabilization is intended to reduce intermittent holes and unstable inpaint decisions around silhouettes and fine detail.
- History stores only a compact confidence/state byte. It never stores color, depth, motion vectors or reconstructed frames; the color reconstruction kernel continues to use current-frame data only.
- Coordinate and non-finite hard rejects remain immediate. Soft ambiguity can receive a single bounded grace observation; repeated drops are handled immediately, with limited confidence recovery and conservative rearming.
- NVIDIA's native decision remains a lower bound: the addon cannot reject work accepted by the native decision.
Safety and compatibility
- Temporal inpaint requires exact kernel, module, ABI, stream and phase-sequence validation. Missing or ambiguous sequencing falls back to Local V3.
- Geometry and inpaint share a 96 MiB maximum temporal-memory budget. At 4K/6x their combined allocation is 99,532,800 bytes, approximately 94.9 MiB.
- If the combined history does not fit, temporal inpaint falls back first, preserving temporal geometry where possible.
- Inpaint fallback order: Temporal V3 → Local V3 → V2 → V1 → native.
- V3 provider patching remains restricted to exactly validated DLSS-G 310.9.0 and 310.9.1. Unknown providers do not receive V3 patches.
- Warp, photometric thresholds, geometry cubins and Reflex/Pacing behavior from 1.3.0 remain unchanged. Pacing overrides remain opt-in.
CUDA dispatch and diagnostics
- Extended the atomic CUDA dispatch table to recognize both geometry and inpaint kernels. Once association is complete, unrelated launches bypass state locks and CUDA name/module/ABI queries.
- Added requested/effective inpaint mode, installed cubin, history-memory usage, resets and fallback diagnostics.
Validation and limitations
- All 21 automated tests passed; the Release build and quality PTX JIT validation completed successfully.
- Provider probes selected Temporal V3 replacements for both 310.9.0 and 310.9.1. These probes validate patch installation, not visual execution in every game.
- Local and Temporal inpaint decision cubins use 48 registers and 784 bytes of shared memory. Their code sections are 9,344 and 10,368 bytes respectively.
- This release does not claim completed cross-game image-quality, latency or performance benchmarking.
- The older Witcher 3 vegetation/hair cadence artifact remains under investigation; no speculative fix for that issue is included.
- No complete NVIDIA DLL or provider package is included in the download.
Installation
Replace the previous renodx-mfgunlock.addon64 with the attached file. Keep only one MFG Unlock addon installed, retain your existing configuration, and fully restart the game.
Existing project credits and the MIT license are preserved. See the repository README for the original addon, kernel research and contributor acknowledgements.
SHA-256 (renodx-mfgunlock.addon64): 9A42464ACCC3191B86885CE2D747246C43636225153D99A1D60113453836F68E