Safeguards for a full graphics card, and a recorder that records what you see. From a player's report (a 4090 at 40 to 50 fps: even with Lossless Scaling's frame generation alone, uneven with an addon, a one-frame warp of the whole picture with
Neural Rendering) and from the owner's flicker: two guesses at the cause are guarded against, the real output can now be recorded for looking into it, and the addons say how much video memory they hold. Nothing changes a picture unless the card is held up.
- Two safeguards against the catch-up hitches a player described with Neural Rendering or the DLSS Upscaler on a full card (a 4090 at 40 to 50 fps: Lossless Scaling's frame generation alone was even; with DLSS "uneven, catching up", with Neural
Rendering added "fine for about 120 frames, then one or two frames of 100 ms with the whole picture warped"; G-Sync, so not the display). (1) The upscalers wait on the GPU for their picture rather than show one twice, which holds Lossless
Scaling's queue up with the upscaler's: when the game holds the card, one very late picture is a hitch of that length and the frames behind it catch up. A picture that takes over 50 ms now switches the GPU wait off for 10 s (a picture
is repeated instead, a hitch of one frame), and says so in the log; the periodic line counts the slow pictures and the slowest. (2) Neural Rendering moved a model result that arrived many frames late along the motion by all of them, which
warps the picture; a result more than three frames old (plus one for each frame auto quality skips) is now left off that picture, and the log says so. Both are guesses from the description, to be checked against their logs and traces. - The addons log how much video memory they hold. The owner saw Lossless Scaling hold over a gigabyte more than it had been (15.2 to 14.1 GB when it closed). Each addon now logs the process's video memory (DXGI's own figure)
before its engine starts, after it, and with the periodic lines ("video memory: this process holds N MB on the card ... N MB since the engine started").nr_srevalandnr_nrevalprint what their engine adds. Measured offline at 2562x1442
in HDR: the upscaler (DLSS, with its textures) adds 281 to 372 MB (sharpening and steady sharpening on, 1.5x or 1:1), Neural Rendering 298 to 412 MB (model at 25 % to 50 %); at 4K the frame-size parts are 2.25 times that, and the
upscaler's shared textures with Lossless Scaling (two frames, three pictures, 66 MB each at 4K in HDR) come on top: with both addons on, over a gigabyte is what they hold. - "Record what is shown" now works for the upscalers with Lossless Scaling's frame generation. The setting said "(with frame generation of our own)" and did nothing else: the recordings stayed the frame going to the
upscaler, so a flicker seen on screen could not be looked into (the recordings of 2026-10-01 were the inputs, and showed no flicker the addons add). With it on, the upscaler addons record the output swap chain's back
buffer at Present, as it goes to the screen (after the upscaler, the generated frames included, tagged), in place of the upscaler's input. The recording can be scored withnr_sreval's no-reference flicker, or toggled
between the enhanced and the plain picture (F7) in the same scene to see what the chain adds.