cs.CVOct 11, 2026

Frozen History: When Temporal Losses Stop Accumulation in Learned Real-Time Supersampling

Authors: Hao Yu, Guoying Zhao

Organizations: ELLIS Institute Finland · CMVS, University of Oulu, Finland

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Learned real-time supersampling accumulates jittered low-resolution samples along motion vectors. Many methods use a temporal consistency loss to reduce flicker. We show that with accurate registration, this loss can suppress accumulation: on static content, the network copies its warped history at almost no temporal cost. Camera jitter in motion vectors raises this cost and can hide the effect. We identify four registration errors that training can conceal: jitter-carrying motion vectors, one motion field for both resolution lattices, jittered depth, and an unregistered skip path. A model-free oracle shows that correct registration makes freezing two orders of magnitude cheaper on static pixels. Controlled training shows that registration reverses the temporal weight's effect on PSNR. Under exact registration, increasing the weight reduces pixel variance at the expense of accuracy and perceived quality, with a 0.7 to 0.9 dB loss on static frames at weight 0.25. This reversal recurs across three network families and two scales; the loss also reduces accuracy on unjittered STSS data. With a small temporal weight and a normalised sample accumulator, a plain registered network improves the best published QRISP PSNR by 0.34 to 0.62 dB at ×\times2, ×\times3 and ×\times4 upscaling to 1080p. At ×\times2, its static pixel standard deviation is one quarter that of DLSS 2. Without retraining, it outperforms classical temporal upsampling by at least 4.4 dB on Arm's jittered Bistro capture. Our most accurate ×\times4 model runs in 2.03 ms per 1080p frame on an H200. Code: https://github.com/CV-AC/Frozen_history

Figures & tables

Figure 1: Registered supersampling. (a) The low-resolution trunk predicts colour, reliability and retention for the normalised output accumulator; the frozen-output limit assumes locally constant accumulated weight. (b) A static reference has small copy cost under R; a large temporal weight then suppresses new samples. (c) The output and sample lattices require different motion fields. (d) Model-free copy cost over chained reprojections. Section 4 details the registration corrections; Section 5.1 derives the copy cost.
Figure 2: Registration and temporal weight. (a,b) Test-set PSNR and static JOD against static pixel STD for λT∈{0,0.05,0.25,1}\lambda_{T}\in\{0,0.05,0.25,1\} ; arrows indicate increasing weight. Filled markers: our ×\times 4 network, 45k iterations; hollow diamonds: QRISP-style network, 15k. (c,d) PSNR change and static copy fraction when the weight rises from 0 to 0.25, across architectures, scales and loss forms. R uses full registration; J uses the original misregistered conventions. Under R, the warped ℓ1\ell_{1} term increases copying while reducing accuracy and static JOD. Grey rows test unwarped and squared terms, including larger weights. Section 5.4 defines copy fraction; Supplementary Tables S2 – S5 give full results.
 J: Δ\Delta PSNR-YR: Δ\Delta PSNR-Y
Settingstaticmovingallstaticmovingall
Ours, ×\times 4+0.11+0.10+0.09−- 0.86−- 0.24−- 0.27
Ours, ×\times 2+0.23+0.20+0.20−- 0.68−- 0.05−- 0.09
QRISP-style a+0.31+0.23+0.23−- 0.85−- 0.30−- 0.32
Direct prediction+0.12+0.12+0.11−- 0.70−- 0.29−- 0.31
Ours, λT=0.05\lambda_{T}{=}0.05+0.05+0.07+0.07−- 0.190.00−- 0.02
Table 1: PSNR-Y change from λT=0\lambda_{T}=0 to 0.25 (last row: 0.05), relative to the same network and pipeline. The warped ℓ1\ell_{1} term costs more accuracy on static than moving frames under R.
MethodIter.ms  ↓\,\downarrowPSNR-Y  ↑\,\uparrowSSIM-Y  ↑\,\uparrowLPIPS  ↓\,\downarrowFLIP  ↓\,\downarrowJOD  ↑\,\uparrowtPSNR 60↑{}_{60}\uparrowStatic PSNR  ↑\,\uparrowStatic FLIP  ↓\,\downarrowStatic stdJOD s
×\times 2 (960 ×\times 540 →\rightarrow 1920 ×\times 1080)
Ours (192 ×\times 8)100k—32.690.92460.0910.05708.6133.5737.570.04580.0968.26
DLSS 2 (shipped frames)——30.120.88210.1510.07328.2231.5536.250.05460.417.96
TAAU (classical)——29.990.87400.1620.07437.9931.0935.430.05680.677.61
FSR 1.2 [ 1 ] (shipped, spatial) ¶——29.440.86680.1600.07087.8828.4932.800.05850.0107.08
Bicubic (our harness)——29.310.86490.2170.07197.8128.4032.630.05930.0117.01
Table 2: QRISP quality and run time. † Published QRISP/NSRR scores; other metrics use our harness. Static PSNR and STD use the 18-frame School stop; static FLIP uses 97 static-camera frames; JOD s averages three stops. ms: one recurrent ×\times 4 step on an H200, excluding rendering. Metrics and evaluation populations are defined in Section 7.1 . Dashes indicate unavailable metrics or timings.
Figure 3: Reconstructions at ×\times 2, ×\times 3 and ×\times 4, with crop PSNR-Y and F LIP. Crops maximise our PSNR gain within detailed regions ( ×\times 2, ×\times 4 School: luma ≥\geq 100), against DLSS 2 at ×\times 2, the best baseline at ×\times 3 and ×\times 4 School, and bicubic at ×\times 4 SeaPort. Table 2 reports test-set scores. At ×\times 3 and ×\times 4 the third column is the nearest-neighbour LR input because no DLSS frames are supplied.
Figure 4: Temporal behaviour at ×\times 2. Top: School frame 283 and static STD maps over frames 275–292; right: full-frame STD distributions. Our mean STD is 0.096 versus DLSS 2’s 0.41, with static PSNR 37.57 versus 36.25 dB. Our 90th-percentile STD is 0.24 versus 0.97; 75% of pixels are constant, versus 24% for DLSS 2 and 99.5% for the reference. Bottom: a SeaPort x–t slice across a pan–stop–pan interval. The plotted trajectories show no visible lag; Figure 5 c gives per-frame scores.
Figure 5: Whole-sequence and per-frame evidence at ×\times 2. (a) PSNR-Y and (b) JOD on each test sequence; labels give our margin over the strongest plotted baseline. (c) SeaPort PSNR-Y, with camera stops shaded. Our model leads in PSNR on every SeaPort frame and on 1455 of the 1460 scored QRISP frames. School and SpaceShip abbreviate AbandonedSchool and SpaceShipDemo.
(a) Arm Bistro, jittered, zero-shot
MethodPSNR-Y  ↑\,\uparrowSSIM-Y  ↑\,\uparrowLPIPS  ↓\,\downarrowFLIP  ↓\,\downarrowJOD  ↑\,\uparrow
Bicubic22.570.77430.2690.15767.22
Bicubic, registered24.210.82830.2590.12417.70
Lanczos-322.410.76620.2910.15757.17
TAAU (classical)27.160.90750.1110.09228.58
TAAU, clamp (classical)27.230.91030.1070.09228.61
Table 3: Other benchmarks at ×\times 2. (a) Arm Bistro, zero-shot from QRISP, with matched 45k history controls. (b) STSS, retrained on one Lewis file and scored with the authors’ code. Temporal weight reduces PSNR on all four STSS scenes.
VariantSeedPSNR-Y ↑\uparrowLPIPS ↓\downarrowSTD s
(a) Ours, ×\times 4, 45k, λT=0.05\lambda_{T}=0.05
Normalised, 30 frames029.740.1960.43
Single predicted blend029.710.1950.53
Normalised, 16 frames029.600.2080.33
(b) History, 45k, λT=0.05\lambda_{T}=0.05 , 30 frames
×\times 2, recurrent—32.580.0950.11
Table 4: Accumulation, history, registration and temporal-weight ablations. Panel (c) removes individual registration corrections; the other panels use full registration. Panel (d) averages two seeds.
Figure S1: Model-free copy cost. PSNR after carrying a frame forward by kk chained reprojections: the high-resolution reference against the target reference on the 1080p output lattice (a), and the low-resolution render against the target render on the sample lattice (b), with registered (blue) and misregistered (orange) motion (jitter-carrying on the output lattice, unjittered on the sample lattice), for moving (solid) and static (dashed) camera frames. All hold lengths use a common set of valid pixels. No network is involved. On the output lattice, a one-frame copy reaches 48.7 dB on the static-camera frames with registered motion, compared with 24.2 dB with jitter-carrying motion. On the sample lattice the shipped jitter-carrying motion is the correct one.
RequirementMeasurementValue
(1) history on the output latticecopy cost ( k=1k=1 , ℓ1\ell_{1} ) on static pixels, mm vs. m~\tilde{m}3.2×10−43.2\times 10^{-4} vs. 3.76×10−23.76\times 10^{-2} (118 ×\times )
 same on the longest camera stop / on moving frames1304 ×\times / 2.2 ×\times
 held-copy PSNR ( k=1k=1 , all frames), mm vs. m~\tilde{m} , ×\times 4 / ×\times 3 / ×\times 235.71/25.45, 35.89/26.51, 36.09/28.30 dB
(2) state on the input latticeprevious LR frame carried with m~\tilde{m} (Equation 5 ) vs. with mm28.35 vs. 26.52 dB
(3) unjittered depthmean ∣| reprojected depth change ∣| , static segment, jittered vs. unjitteredabout 103×10^{3}\times
(4) registered skipgain of the registered bicubic skip, moving SeaPort frames (5 frames)+0.86+0.86 to +1.53+1.53 dB
Table S1: Model-free measurements quoted in Sections 4 and 5.2 of the paper (QRISP test set, ×\times 4 unless noted; no network involved).
λT\lambda_{T}PSNR-Y ↑\uparrowtPSNR 60 ⁣↑{}_{60}\!\uparrowStd sJOD s ⁣↑{}_{\mathrm{s}}\!\uparrowα~s\tilde{\alpha}_{\mathrm{s}}FLIP ↓\downarrow
R: registered pipeline
029.7530.590.777.640.0340.0801
0.0529.7430.950.437.552.3×10−32.3{\times}10^{-3}0.0829
0.2529.4831.210.177.263.2×10−53.2{\times}10^{-5}0.0881
128.9631.320.0866.709.6×10−79.6{\times}10^{-7}0.0980
J: misregistered pipeline
Table S2: Registration ×\times temporal weight: our 192×8192\times 8 network at ×\times 4, trained for 45k iterations. Only the registration pipeline and λT\lambda_{T} differ. Static standard deviation and sample share use the AbandonedSchool stop; static JOD averages the three static segments. Temporal PSNR uses a shared unjittered motion field. Metric definitions and averaging are given below.
Figure S2: The static AbandonedSchool stop ( ×\times 4) for the R models with λT=0\lambda_{T}=0 , 0.250.25 and 11 and the J model with λT=0.25\lambda_{T}=0.25 : output crops at frame 283, temporal standard-deviation maps over frames 275–292 (8-bit intensity units), and per-pixel median deployed new-sample shares α\alpha over the same frames. Numbers below the output crops are crop PSNR-Y and F LIP; static STD (float output) and segment median α\alpha are full-frame segment statistics as in Table S2 .
Figure S3: Training dynamics on the grid of Table S2 . Top: mean new-sample share α\alpha on training crops. Bottom: fraction with ρ≥0.999\rho\geq 0.999 . Left: R; right: J. These training averages are distinct from the deployed static-segment medians.
 Δ\Delta PSNR-YΔ\Delta JOD scopy fraction
SettingJRJRJR
Ours+0.09−- 0.27+0.03−- 0.380.20 →\rightarrow 0.230.85 →\rightarrow 0.99
Ours, ×\times 2+0.20−- 0.09+0.01−- 0.140.32 →\rightarrow 0.450.87 →\rightarrow 1.00
QRISP-style a+0.23−- 0.32+0.10−- 0.690.11 →\rightarrow 0.160.70 →\rightarrow 0.98
Direct prediction+0.11−- 0.31+0.01−- 0.340.21 →\rightarrow 0.310.21 →\rightarrow 0.77
Unwarped term+0.08+0.09−- 0.11−- 0.100.20 →\rightarrow 0.550.85 →\rightarrow 0.96
Table S3: Architecture, scale and loss comparisons corresponding to Figure 2 c,d of the paper. Changes are from λT=0\lambda_{T}=0 to 0.250.25 under J and R. The first four rows use the warped ℓ1\ell_{1} term; the remaining rows test other loss forms and unjittered STSS input.
Run PSNR-YSSIM-YLPIPStPSNR 60Std sJODJOD sα~s\tilde{\alpha}_{\mathrm{s}}copy
Our network, ×\times 4
RλT=0\lambda_{T}{=}029.750.85600.17730.590.7697.717.640.0340.855
RλT=0.25\lambda_{T}{=}0.2529.480.84320.22131.210.1747.607.263.2×10−53.2\times 10^{-5}0.986
JλT=0\lambda_{T}{=}029.720.85280.19330.151.3617.707.580.610.199
JλT=0.25\lambda_{T}{=}0.2529.810.85710.17930.281.3327.717.610.460.234
Our network, ×\times 2
Table S4: Every run behind Figure 2 c,d of the paper (R: registered pipeline, J: misregistered pipeline). For the warped ℓ1\ell_{1} term, raising λT\lambda_{T} from 0 to 0.25 increases the copy fraction under R in all tested architectures and at both scales; the increase is smaller under J. The final block compares alternative loss forms.
  JR
SettingλT\lambda_{T}static (97)moving (1363)allstatic (97)moving (1363)all
Ours, ×\times 40→0.250\rightarrow 0.25+0.11+0.10+0.09−- 0.86−- 0.24−- 0.27
Ours, ×\times 40→0.050\rightarrow 0.05+0.05+0.07+0.07−- 0.190.00−- 0.02
Ours, ×\times 40→10\rightarrow 1+0.02+0.04+0.02−- 1.98−- 0.74−- 0.80
Ours, ×\times 20→0.250\rightarrow 0.25+0.23+0.20+0.20−- 0.68−- 0.05−- 0.09
Ours, ×\times 2 a0→0.050\rightarrow 0.05———−- 0.20+0.04+0.02
Table S5: Change in PSNR-Y from zero temporal weight to the indicated weight under J and R, reported separately for static-camera frames, moving frames and the complete test set.
(a) QRISP-style architecture, QRISP, ×\times 4, 15k
PipelineλT\lambda_{T}PSNR-Y ↑\uparrowtPSNR 60 ⁣↑{}_{60}\!\uparrowStd sJOD s ⁣↑{}_{\mathrm{s}}\!\uparrowα~s\tilde{\alpha}_{\mathrm{s}}
R028.7329.611.257.280.031
R0.2528.4130.340.226.591.1×10−51.1{\times}10^{-5}
J028.3228.821.986.920.42
J0.2528.5429.031.867.020.31
(b) STSS Arena, ×\times 2, no jitter, 45k
Table S6: QRISP-style architecture and STSS Arena. (a) Single- α\alpha blend, ×\times 4, 15k iterations, with our pipeline and schedule. (b) Arena at ×\times 2 without jitter. These are mechanism comparisons; the QRISP-style scores do not reproduce the published QRISP protocol.
V1. Jitter: ∣| LR −- reference at the sample positions ∣|
convention used0.0248
yy mirrored0.0455
xx mirrored0.0415
axes swapped0.0433
zero jitter0.0385
V2. Motion: ∣| warped previous reference −- reference ∣|
Table S7: Model-free checks of jitter sign, motion convention and the two-lattice correction on the Arm capture. The convention used gives the lowest error among the tested alternatives in each check.
RunPSNRSSIMLPIPSFLIPJODtPSNR
R, λT=0\lambda_{T}{=}0 ( ×\times 2 grid)31.120.94760.0410.07559.2232.94
R, λT=0.25\lambda_{T}{=}0.2530.960.94900.0440.08039.2033.78
J, λT=0\lambda_{T}{=}0 , own conv.31.060.94910.0400.08459.1232.10
J, λT=0.25\lambda_{T}{=}0.25 , own conv.31.320.95170.0380.08148.9932.84
J, λT=0\lambda_{T}{=}0 , jitter-free motion23.930.80810.1700.15037.2924.28
Ours, sRGB space30.530.95650.042—9.1032.35
Table S8: Additional Arm Bistro results at ×\times 2, zero-shot. R and J use their own conventions unless noted. Wrong-convention runs lose 6.1 to 12.0 dB relative to the same checkpoint with its own conventions. The checkpoint and input changes are specified below.
Figure S4: Selected ×\times 2 crops on the Arm Bistro capture, zero-shot (top), and STSS Lewis with our STSS-trained model at λT=0.05\lambda_{T}=0.05 (bottom): bicubic, TAAU, our model and the reference. Full reference frames are shown at the left; labels give crop PSNR-Y and F LIP.
 ArenaLewisSunTempleSubway
 PSNR-YFLIPJODPSNR-YFLIPJODPSNR-YFLIPJODPSNR-YFLIPJOD
Bicubic33.240.05618.9332.830.06658.2531.420.08047.9329.870.07007.36
TAAU32.180.07148.8332.900.06988.3131.120.08627.8630.390.07127.54
TAAU, clamp32.110.07248.8132.810.07048.2931.020.08737.8330.370.07197.54
Ours, λT=0\lambda_{T}{=}034.160.07108.9937.530.04669.1631.640.08667.9831.080.07037.65
Ours, λT=0.05\lambda_{T}{=}0.0533.610.08758.8637.620.04769.1731.610.08657.9331.090.07107.68
Table S9: STSS ( ×\times 2, no jitter) on our protocol, per test scene. Compared with λT=0\lambda_{T}=0 , λT=0.25\lambda_{T}=0.25 lowers PSNR on all four scenes, with the largest drop of 3.70 dB on Arena.
VariantPSNR-Y ↑\uparrowΔ\DeltaLPIPS ↓\downarrowtPSNR 60 ⁣↑{}_{60}\!\uparrowStd s
(a) R, λT=0.05\lambda_{T}=0.05 , 192 ×\times 8, ×\times 4, 45k
final recipe29.74—0.19630.950.43
single- α\alpha blend accumulator29.71−- 0.020.19530.950.53
16-frame training clips29.60−- 0.140.20830.950.33
(b) R minus one piece, λT=0\lambda_{T}=0 , 192 ×\times 8, ×\times 4, 15k
control, all pieces29.31—0.20330.190.83
Table S10: Ablations. (a) Components of the final recipe at 45k iterations. (b) One registration component removed at a time, with λT=0\lambda_{T}=0 and 15k iterations. Each difference uses the control with the same seed; the failed seed-0 depth run is retained in Table S11 .
VariantPSNR-Y ↑\uparrowΔ\DeltaLPIPS ↓\downarrowtPSNR 60 ⁣↑{}_{60}\!\uparrowStd s
control, all pieces29.31—0.20330.190.83
control, all pieces, seed 129.38—0.22330.250.74
−- sample phase29.22−- 0.090.20830.080.81
−- registered skip29.28−- 0.030.20530.140.80
−- LR re-jitter29.19−- 0.120.21030.120.81
−- unjittered depthdiv.div.div.div.div.
Table S11: All registration leave-one-out runs: R, λT=0\lambda_{T}=0 , 192×8192\times 8 , ×\times 4, 15k iterations. This includes the failed seed-0 depth run and complements Table S10 .
Figure S5: Selected crops at ×\times 3 and ×\times 4: bicubic, our classical TAAU, our model and the reference. The full reference frames mark the crop locations, and labels give crop PSNR-Y and FLIP. The ×\times 4 crop lies in the static segment. Selection is described in the text and in Section S16 .
ModelSizeIter.ParamsGFLOPH200 msPSNR-Y  ↑\,\uparrow
Measurements from main Table 2
QRISP-M †f32 / m3500k71k180.44329.06
QRISP-L †f64 / m5500k0.27M700.69829.42
Ours, shift-MLP64 ×\times 6200k0.17M380.20129.32
Ours, shift-MLP64 ×\times 8200k0.20M470.30129.35
Ours64 ×\times 4200k0.36M890.33229.32
Table S12: Recurrent-step latency at ×\times 4, 1080p output and batch 1, excluding rendering. H200 measurements are identical to the main paper. Daggered PSNR values are published QRISP results.
Figure S6: PSNR-Y and H200 recurrent-step latency at ×\times 4, from Table S12 . Circles: our convolutional trunks; squares: our shift-MLP trunks, both trained for 200k iterations. Diamonds: published QRISP PSNR paired with our re-implementation’s latency. Lines connect configurations within each method.
   Δ\Delta PSNR-Y (dB)
ModelGPUIter.fp16int8int8, worst scene
64 ×\times 4A10015k0.000−- 0.128−- 0.214
64 ×\times 4H20015k0.000−- 0.127−- 0.214
192 ×\times 8A10045k0.000−- 0.352−- 0.489
192 ×\times 8H20045k0.000−- 0.341−- 0.480
Table S13: PSNR cost of TensorRT int8 inference ( ×\times 4): change in PSNR-Y against the float32 model over one unbroken 300-frame pass of each scene.
ModelH200 ms60 Hz120 HzBreak-even ∗
QRISP-M0.4433%5%2 116 fps
QRISP-L0.6984%8%1 343 fps
Ours, shift-MLP 64×664\times 60.2011%2%4 664 fps
Ours, shift-MLP 64×864\times 80.3012%4%3 115 fps
Ours 64×464\times 40.3322%4%2 824 fps
Ours 96×496\times 40.5243%6%1 789 fps
Table S14: Current H200 run time as a share of a frame interval ( ×\times 4, 1080p output). Percentages and break-even thresholds are derived from Table S12 , with exact intervals 1000/601000/60 and 1000/1201000/120 ms.
Iter.λT\lambda_{T}SeedPSNR-Y  ↑\,\uparrowLPIPS  ↓\,\downarrowStatic PSNR  ↑\,\uparrowStatic stdMed. α\alpha
15k0028.730.24133.591.300.031
15k0.05028.690.26133.530.641.2 ⋅10−3\cdot 10^{-3}
15k0.25028.410.31332.660.231.1 ⋅10−5\cdot 10^{-5}
45k0029.050.21033.891.160.024
45k0129.040.21333.891.120.024
45k0.05028.970.23333.710.596.9 ⋅10−4\cdot 10^{-4}
Table S15: The QRISP-style network of Section 5.4 of the paper (QRISP-L’s architecture) trained with our registered pipeline. A small temporal weight nearly halves the static pixel standard deviation at 0.06 dB lower PSNR (45k iterations, two seeds); a large one freezes the accumulator and costs 0.9 dB of static PSNR (15k). The scores are not comparable with the published ones (500k iterations, their pipeline).
TierTrunkSizelrPSNR-Y  ↑\,\uparrowΔ\Delta PSNR
(a) Conventional trunks
Splain conv (control)64 ×\times 46.0e-428.75—
 ECB re-parameterised64 ×\times 46.0e-428.98+0.24
 plain conv64 ×\times 42.0e-328.98+0.23
 ECB re-parameterised64 ×\times 42.0e-328.88+0.13
 U-Net 1-2-164/1286.0e-428.76+0.01
Table S16: Trunk alternatives and learning rates: ×\times 4, 15k iterations, registered pipeline and the same training recipe. Differences use the first control of the same tier, with the medium-tier seed-1 ECB compared against its seed-1 control. Only reconstruction accuracy is reported.
Figure S7: Failure case: particle sprites without motion vectors (SpaceShipDemo/0001, ×\times 4; the R and J grid models at λT=0.05\lambda_{T}=0.05 , Table S19 gives both weights). Both models smear the particle sprites. The R model has larger error in the shown crop (22.1 versus 23.2 dB crop PSNR-Y; the labels give crop PSNR-Y and F LIP).
MethodPSNR-Y  ↑\,\uparrowΔ\DeltaSSIM-Y  ↑\,\uparrowLPIPS  ↓\,\downarrowFLIP  ↓\,\downarrowJOD  ↑\,\uparrowStatic std
(a) Stateless and classical temporal upsamplers, QRISP
×\times 2
Bicubic29.31 0.86490.2170.07197.810.011
Lanczos-329.17−- 0.140.86220.2330.07187.780.011
Lanczos-3, jittered input27.95−- 1.360.82710.237———
TAAU (variance clipping)29.99 0.87400.1620.07437.990.67
Table S17: Additional rows. (a) Stateless and classical temporal upsamplers on QRISP ( Δ\Delta : against bicubic upsampling; for the clamp rows, against the TAAU with variance clipping). (b) Our network with and without history at equal training length (45k iterations), on QRISP and zero-shot on the Arm capture.
 PSNR-Y  ↑\,\uparrowF LIP  ↓\,\downarrowJOD  ↑\,\uparrowpixel std (8-bit)
Sequence / segmentOursDLSS 2OursDLSS 2OursDLSS 2OursDLSS 2Ref.
whole sequences (frames 8–299)
AbandonedSchool/000025.8724.580.11370.12827.777.35———
AbandonedSchool/000137.9235.690.03840.04869.148.80———
SeaPort/000030.0427.390.05250.06948.407.92———
SpaceShipDemo/000036.8333.270.04000.05339.238.80———
Table S18: Ours and DLSS 2 at ×\times 2 on every test sequence and every static segment. Ours is higher in PSNR-Y and JOD and lower in F LIP on all five sequences, and higher in JOD, lower in F LIP and lower in pixel standard deviation on all three static segments.
  Abandoned- SpaceShipDemo 
λT\lambda_{T} SchoolSeaPortall00000001Mean
0R28.9527.5932.7233.9631.4729.75
 J28.8627.4432.8533.8731.8329.72
 R −- J+0.10+0.15−- 0.13+0.09−- 0.35+0.04
0.05R28.9427.5532.7233.8131.6329.74
 J28.9127.4932.9533.9931.9129.78
Table S19: Per-scene PSNR-Y of the R and J ×\times 4 grid models (192 ×\times 8, 45k iterations), with the two SpaceShipDemo sequences.
Methodstatic (97)moving (1363)all
Ours (192 ×\times 8, 100k)0.04580.05870.0570
DLSS 2 (shipped frames)0.05460.07540.0732
TAAU (classical)0.05680.07600.0743
Bicubic0.05930.07310.0719
Table S20: F LIP on static-camera and moving frames ( ×\times 2, QRISP test set).
TrunkλT\lambda_{T}ClipSeed 0Seed 1∣Δ∣|\Delta|
96 ×\times 80.053029.16429.1680.003
192 ×\times 80.251628.81828.8270.009
192 ×\times 80.051629.27729.2660.011
192 ×\times 8 b0.053029.42429.3980.026
96 ×\times 8, ECB0.053029.37529.3460.028
192 ×\times 801629.31229.3760.064
Table S21: Seed pairs: every pair of 15k-iteration runs ( ×\times 4) whose configurations differ only in the seed.
Figure S8: Per-frame PSNR-Y (top) and FLIP (middle) on SeaPort/0000 at ×\times 2, scored frames 8–299: ours, DLSS 2, TAAU and bicubic. Grey bands mark camera stops; the bottom strip shows median camera motion. The legend gives sequence means.
Figure S9: Gallery at ×\times 2 (selected crops). Each row shows the full reference frame with the crop box, bicubic upsampling, our classical TAAU, DLSS 2 (shipped frames), our model and the reference; below each crop is its F LIP error map (0 to 1), labelled with crop PSNR-Y and F LIP. Our model recovers thin and periodic structure (branches, lettering, floor gratings) where the baselines alias, blur or break it up; in the SeaPort frame-129 crop, TAAU produces fragmented lettering and scores 14.7 dB PSNR-Y.
Figure S10: Gallery at ×\times 3 (selected crops), as in Figure S9 ; the QRISP dataset provides DLSS 2 frames only at ×\times 2.
Figure S11: Gallery at ×\times 4 (selected crops), as in Figure S9 . At ×\times 4, the input has one low-resolution pixel per 16 output pixels. In these selected crops, our model reconstructs grass blades, branches and floor gratings with less blocking and moiré than bicubic upsampling and TAAU.
More in this topic

Image Super-Resolution

Image super-resolution reconstructs high-resolution images from lower-resolution observations, balancing detail, fidelity, and computational cost.

See all 248 papers

Explore similar work

CardsList

  1. Frozen Flows Forget: Diagnosing and Restoring Lost Motion in a Latent-flow World Model

    Sep 23, 2026Xiwen Chen, Rigaudiere Z. Li, Zhiruo Zhou +2World Model LearningLatent Dynamics Modeling

    Shenzhen International Graduate School, Tsinghua University · Shanghai Jiaotong University · Wuhan University of Technology

  2. RelayVSR: Large-Small Model Collaboration for Efficient Real-World Video Super-Resolution

    Sep 29, 2026Xijun Wang, Xin Li, Zirui Lang +3Video Super-Resolution

    University of Science and Technology of China

  3. TVRN: Invertible Neural Networks for Compression-Aware Temporal Video Rescaling

    May 15, 2026Xinmin Feng, Li Li, Dong Liu +1Invertible Neural NetworksVideo Compression

    MOE Key Laboratory of Brain-Inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei 230093, China