Choose this Enlargement mode below 100% model resolution for ordinary post-upscale or finished-picture NR. Keep both early-generation options off. INI: Transfer=2; spatial Matched residual remains the default (1). At 100%, ordinary full-resolution NR replaces the private context.
Lighting + colour (Transfer=3) brings the Display Filter structural composition path into OptiScaler. It measures lighting gain and chromaticity changes against the exact low-resolution model input, resizes those separately with bilinear filtering, and reconstructs the answer on the full-resolution proxy. Existing HDR composition, strength and skin controls then apply once. Native Vulkan and DX12 processing share this shader path, including pre-upscale and finished-picture placement. The resolution here is the NR injection surface, not necessarily the swapchain size.
Lighting + colour + DLSS (Transfer=4) sends a bounded carrier of the same fields through the existing private DLSS SR feature. It has the same post-upscale DX12 requirement as Transfer=2. Changing between the two DLSS modes retires the old feature before creating a new one, so history cannot mix RGB residuals and lighting/chromaticity fields. The unscaled field also provides a spatial fallback when DLSS alone would cross the HDR inverse curve's white limit.
These modes are opt-in. Values 0, 1 and 2 retain their previous algorithms, and native/supersampled NR bypasses the new reconstruction. Raw model inspection still shows the actual NR answer. Fully black source pixels remain black: relative lighting cannot reproduce newly generated light at true zero. Near-black gain uses paired neighbouring samples. This can reduce halos introduced by resizing; it cannot remove halos already generated by the model. Game verification remains necessary.
NR runs on a downsampled reconstructed picture. Encode its linear-proxy difference from that same input around neutral 0.5, enlarge with independent DLSS SR, then add it to the full-resolution original. Existing HDR, strength and skin controls compose the matched pair. Replace modes decode the reconstructed proxy; model debug view shows the actual low-resolution answer.
Encoding/decoding use a 1/64 linear-light scale to preserve small edits in FP16. This removed dense shadow speckles observed in KCD2 without changing model strength.
Depth/motion are resampled from their active regions; motion becomes pixels at model resolution. Jitter is zero because game reconstruction has already occurred. Finished-picture processing uses matching captured guides/metadata; hold zeroes velocity. The private context owns history/parameters, uses unit exposure and disables sharpening/auto-exposure. This route never invokes private RR.
The actual command-list submission identifies the producer queue; a Streamline presentation-queue hint must not invalidate it. Resize, authoritative queue changes and Retry retire the context through its own GPU tracker. Active/retired contexts receive reset/submission notifications; unrelated NR work cannot pin them. Retirement is bounded.
Failure retains the clean frame with a status, without automatic spatial fallback. Native Vulkan and pre-upscale placement are unsupported; D3D12 bridges reuse this implementation. Shader operations 9/10 preserve existing numbers/layout and have matching DX12/Vulkan binaries.
WARP checks cover reconstruction, guide scaling, neutral identity and signed shadow edits through FP16; Vulkan, lifetime/proxy and real DLSS adapter checks also passed. Live checks found a remaining faint post-upscale grid in KCD2 and possibly Hogwarts. Spatial Matched residual avoids that observed artifact. HDR and motion quality remain experimental; private DLSS adds GPU/VRAM cost beyond model timing.