---
kind: concept
slug: image-super-resolution
title: Image super-resolution
summary: >-
  Learned upscaling of a low-resolution image or field, and ways to judge or explain the result. Covers
  ESRT, GRL, MambaIR, StableSR, SeeSR, DiffIR, eDiff-I, BasicVSR, RBPN, ACNN-SR, WIRE, a physics-informed
  diffusion model for flow fields, KernelGAN blur-kernel estimation, LAM attribution and a no-reference
  metric. Protocols differ.
tags: [super-resolution, image-restoration, lightweight-model, interpretability, image-quality-assessment, diffusion, video, medical-imaging, kernel-estimation, semantic-prompts, transformer, fluid-dynamics]
papers: ["1612.05890", "1903.10128", "2011.11036", "2108.11084", "2211.01324", "2211.14680", "2301.05187", "2303.00748", "2303.09472", "2402.15648", "2305.07015", "2311.16518", "2012.02181", "2104.13371", "1705.08302", "1909.06581"]
related: [vision-transformer, self-attention, shifted-window-attention, channel-attention, state-space-model, efficient-vision-backbone, feature-attribution, diffusion-models, latent-diffusion, classifier-free-guidance, generative-prior-inverse-problems, recurrent-neural-network, residual-learning, denoising-autoencoder, semantic-segmentation, structural-similarity, image-quality-assessment, random-forest, generative-adversarial-network, image-deblurring, image-denoising, diffusion-inpainting, optical-flow, text-to-image-generation, physics-informed-neural-network]
status: draft
updated_at: 2026-11-01
---

## Overview

Super-resolution (SR) takes a low-resolution (LR) image and predicts a
high-resolution (HR) one, so the network has to produce detail that the input
does not contain. The cards in this wiki cover several questions: how a
reconstruction network gathers context from many pixels without prohibitive
cost, how a frozen diffusion model can restore real photos with unknown
degradations [arxiv:2305.07015 §3.1], how semantic prompts from a degradation-aware
tagger can steer such a model [arxiv:2311.16518 §3.2], how SR appears as one stage of a diffusion
cascade, how information is shared across frames in video [arxiv:2012.02181 §3.1],
and how a second-order version of that video design aligns frames with flow-guided
deformable offsets [arxiv:2104.13371 §3.1] [arxiv:2104.13371 §3.2], how a
training-only shape term can keep medical SR outputs anatomically plausible
[arxiv:1705.08302 §2.4], how a network fitted directly to the measurements of one
image or a few frames can recover detail [arxiv:2301.05187 §4.2]
[arxiv:2301.05187 app. A.4], how to estimate the blur kernel of one unknown-blur
image from that image alone [arxiv:1909.06581 §2], how to see which input pixels
an SR network relies on, and how to score an SR output when no sharp reference image exists. Two later
cards add a recurrent video design that back-projects per-neighbour estimates
[arxiv:1903.10128 §3.1], and a diffusion model that runs on a short restoration
prior instead of on pixels [arxiv:2303.09472 §4.2]. The two
classic reconstruction cards train with an L1 loss and Adam at an initial
learning rate of 2e-4 at the paper level [arxiv:2108.11084 §4.2]
[arxiv:2402.15648 §3.6] [arxiv:2402.15648 §4.1]. The other trained variants
use different losses, listed under Implementation sketch.
Two further cards widen the set. GRL is a restoration transformer whose layers
run attention along thin stripes through a small set of anchor points, which
keeps long-range attention affordable [arxiv:2303.00748 §3.2] [arxiv:2303.00748 §4].
The physics-informed card applies the same denoising idea to two-dimensional
turbulent flow fields rather than photographs, with the fluid equation as an
optional guide [arxiv:2211.14680 §2.3.1].

The scores on this page are not on one scale. The classic reconstruction cards
report PSNR in dB on classic benchmarks. StableSR reports FID, CLIP-IQA and
MUSIQ on real-world sets, alongside PSNR and SSIM, and it tests on Real-ESRGAN
degradations [arxiv:2305.07015 §4.2]. SeeSR reports on DIV2K-Val, RealSR, DRealSR and
RealLR200, and follows the StableSR protocol for the two real-world sets [arxiv:2311.16518 §4.1]. The video card reports PSNR and SSIM, with
REDS4 scored in RGB and the other sets on the Y channel [arxiv:2012.02181 §4],
and the BasicVSR++ card uses the same convention [arxiv:2104.13371 tab. 1].
The ACNN card reports SSIM and a 1 to 5 opinion score on cardiac MR
[arxiv:1705.08302 tab. 3]. The attribution card reports correlations between its
index and image metrics, and the quality metric reports a rank correlation with
human ratings. The WIRE card gives its super-resolution results as an ordering
and a gain over other activations, not as a benchmark table [arxiv:2301.05187 §4.2]
[arxiv:2301.05187 app. A.4]. RBPN reports PSNR and SSIM on Vimeo-90k, Vid4 and SPMCS-11 [arxiv:1903.10128 tab. 9]
[arxiv:1903.10128 tab. 7] [arxiv:1903.10128 tab. 8]. DiffIR reports PSNR and LPIPS
on five SR sets, but its main table compares only with GAN and diffusion
restorers [arxiv:2303.09472 tab. 2]. The metric card was trained on outputs of nine older
SR methods, and it reports no evaluation on the networks described here
[arxiv:1612.05890 §3] [arxiv:1612.05890 tab. 7]. The KernelGAN card reports
PSNR and SSIM on its own synthetic DIV2KRK benchmark, with blur kernels that
differ per image [arxiv:1909.06581 §5.2] [arxiv:1909.06581 tab. 1]. GRL reports
PSNR on classic ×4 sets [arxiv:2303.00748 tab. 5], and the flow-field card reports
an L2 norm and an equation residual on its own simulated fields [arxiv:2211.14680 tab. 1].
Neither number is an image-quality score on the scale above.

## Variants

### Lightweight CNN-Transformer hybrid (ESRT)

A small convolutional body with three High Preserving Blocks works at half
resolution and carries a high-frequency residual. One Efficient Transformer then
attends over unfolded 3×3 neighbourhoods, and its attention is split into four
segments so each segment attends only within itself [arxiv:2108.11084 §3.1]
[arxiv:2108.11084 §3.3.2]. Author-reported ×4 results: Set5 32.19 dB and
Urban100 26.39 dB with 751K parameters [arxiv:2108.11084 tab. 1]. Card:
[Transformer for Single Image Super-Resolution](/papers/2108.11084).

Training memory is the main claim. The authors measure 4,191M for one
Efficient Transformer against 16,057M for an original Transformer, both on
48×48 patches at batch size 16 [arxiv:2108.11084 tab. 4]. The paper does not
state how that baseline was configured.

### State-space restorer (MambaIR)

The network follows a SwinIR-style three-stage layout. Each block replaces window
attention with a state-space scan over four directions of the feature map,
followed by a convolution bottleneck with channel attention [arxiv:2402.15648 §3.2]
[arxiv:2402.15648 §3.3] [arxiv:2402.15648 §3.5]. Author-reported ×4 results on
classic SR: Set5 33.03 dB and Urban100 27.68 dB for MambaIR, against 32.92 dB and
27.45 dB for SwinIR [arxiv:2402.15648 tab. 3]. The lightweight variant reaches
Urban100 26.52 dB at 924K parameters and 42.3G MACs, where SwinIR-light is at
26.47 dB with 930K parameters and 31.8G MACs [arxiv:2402.15648 tab. 4]. Card:
[MambaIR: A Simple Baseline for Image Restoration with State-Space Model](/papers/2402.15648).

The paper does not state the width, depth or state size of its models; those
values exist only in the authors' code, which the card records.

### Anchored stripe attention for restoration (GRL)

GRL is a restoration transformer that also covers denoising, deblurring, JPEG
artifact removal and demosaicking [arxiv:2303.00748 §5]. Each layer runs three
branches in parallel: window attention for the regional range, anchored stripe
attention for long-range context, and a channel-attention convolution for local
structure [arxiv:2303.00748 §4] [arxiv:2303.00748 fig. 6]. The stripe branch
does not build a full attention map. Its M anchors summarise the keys and values
first, then each query reads the anchors, with a stated cost of O(NMd) time and
O(NM) space [arxiv:2303.00748 eq. 2] [arxiv:2303.00748 §3.2]. Stripes alternate
between horizontal and vertical across layers [arxiv:2303.00748 §3.3].

Author-reported ×4 classical SR, PSNR: GRL-B reaches 33.10 dB on Set5 and 28.53 dB
on Urban100 with 20.20M parameters, against 32.92 dB and 27.45 dB for SwinIR at
11.90M [arxiv:2303.00748 tab. 5]. The smaller models are GRL-S at 3.49M, with
27.90 dB on Urban100, and GRL-T at 0.91M, with 27.15 dB [arxiv:2303.00748 tab. 5].
GRL-S is below that SwinIR on Set5 (32.76 dB against 32.92 dB), so the gain is
not uniform across sets [arxiv:2303.00748 tab. 5]. Card:
[Efficient and Explicit Modelling of Image Hierarchies for Image Restoration](/papers/2303.00748).

The paper gives no training protocol in the main text; it states Adam, an L1
loss and an initial rate of 2e-4 [arxiv:2303.00748 §5]. The released configs
for each task differ from that, and the card's Paper vs code section lists them.

### Diffusion prior for real photos (StableSR)

StableSR keeps a pre-trained Stable Diffusion 2.1-base U-Net and its autoencoder
frozen. It trains only a time-aware encoder and the scale-and-shift layers that
modulate the U-Net's residual blocks [arxiv:2305.07015 §3.1] [arxiv:2305.07015
eq. 1]. The encoder reads the LR image and the diffusion step, so its guidance
changes over the sampling trajectory, and the authors place the strongest
guidance where the signal-to-noise ratio is near 5e-2 [arxiv:2305.07015 §3.1]
[arxiv:2305.07015 fig. 3].

Two optional parts sit on top. A controllable feature wrapping module, weighted
by w between 0 and 1, favours realism near 0 and fidelity near 1; it is trained
after the first stage [arxiv:2305.07015 §3.2] [arxiv:2305.07015 tab. 4]. An
aggregation sampler cuts the latent into overlapping 64×64 patches, blends their
noise predictions with Gaussian weights at each step, and so handles arbitrary
sizes without visible seams [arxiv:2305.07015 §3.3] [arxiv:2305.07015 alg. 1].

Author-reported on DIV2K-Valid, StableSR has the best FID, 24.44 against 26.47
for LDM, but lower PSNR and SSIM than several baselines [arxiv:2305.07015 tab. 1].
On the RealSR set its PSNR is 24.65 dB, below 27.30 dB for the RealSR baseline and
26.38 dB for BSRGAN [arxiv:2305.07015 tab. 1]. Table 3 turns off one part at a time: without pixel
colour correction, RealSR PSNR falls from 24.65 to 22.24 dB [arxiv:2305.07015
tab. 3]. Cost is high: 200 DDPM steps take 15.16 s for a 512×512 output on one
V100, and the 4-step SD-Turbo variant takes 0.83 s [arxiv:2305.07015 tab. 5].
Card: [Exploiting Diffusion Prior for Real-World Image Super-Resolution](/papers/2305.07015).

### Semantic prompts from a degradation-aware tagger (SeeSR)

SeeSR also runs Stable Diffusion for ×4 real-world SR, but it conditions the
model on what the image shows. A RAM tagger is first fine-tuned on degraded
inputs (DAPE), so its outputs on a low-resolution image match its outputs on the
clean original [arxiv:2311.16518 §3.2] [arxiv:2311.16518 eq. 1]. The predicted
tags enter as text through the frozen text encoder. The tagger's representation
embeddings enter through extra cross-attention modules, and a ControlNet branch
reads the LR image [arxiv:2311.16518 §3.2] [arxiv:2311.16518 §3.3]. At inference
the sampler starts from the LR latent noised by the training schedule, not from
pure noise [arxiv:2311.16518 §3.4]. The paper does not give the start step.

Author-reported ×4: SeeSR has the best FID, CLIPIQA and MUSIQ on all four test
sets, and the best LPIPS and DISTS on DIV2K-Val [arxiv:2311.16518 §4.2]
[arxiv:2311.16518 tab. 2]. On DIV2K-Val its FID is 31.93, against 36.57 for
StableSR, the best other method there. Its PSNR and SSIM are below the best
baseline on the three paired sets [arxiv:2311.16518 tab. 2]. The reported
cost is 2283.7M parameters and about 7.24 s per 512×512 output with 50
steps on one V100; these figures come from the supplement, not the main text
[arxiv:2311.16518 §4.1]. Card:
[SeeSR: Towards Semantics-Aware Real-World Image Super-Resolution](/papers/2311.16518).

### Diffusion SR stages (eDiff-I)

In a text-to-image cascade, eDiff-I runs a 64 px base model, then a 256 px and
a 1024 px super-resolution stage [arxiv:2211.01324 §4.2]. Both SR stages are
diffusion denoisers. The 256 px stage has 128 channels, self- and
cross-attention at one resolution, and two residual blocks per level
[arxiv:2211.01324 tab. 3]. The 1024 px stage removes self-attention and keeps
only cross-attention, and it is trained on random 256×256 crops and applied at
1024×1024 [arxiv:2211.01324 tab. 4] [arxiv:2211.01324 app. A]. Sampling for the
SR stage uses a DEIS solver of order 3 over 10 steps [arxiv:2211.01324 tab. 7].
Card: [eDiff-I: Text-to-Image Diffusion Models with an Ensemble of Expert Denoisers](/papers/2211.01324).

The paper reports only end-to-end zero-shot FID for the whole cascade, with no
SR-only quality metric [arxiv:2211.01324 tab. 1]. Its SR stages are one part of
a much larger system trained on about one billion pairs [arxiv:2211.01324 §5],
so the design does not transfer to a standalone SR model without retraining and
measurement.

### Diffusion on a restoration prior (DiffIR)

DiffIR runs the diffusion process on a compact restoration prior, a vector of
256 values, instead of on pixels or feature maps [arxiv:2303.09472 §4.1]
[arxiv:2303.09472 §4.2]. In the first stage an encoder that sees the
ground-truth image gives the prior, and a transformer restorer uses it. In the
second stage a conditional denoiser predicts the prior from the low-quality
image in four steps [arxiv:2303.09472 §5.1]. Both stages are trained with L1
terms: one on the restored image and one on the prior [arxiv:2303.09472 eq. 13].

Author-reported ×4 SR, PSNR on Set14, Urban100, Manga109, General100 and
DIV2K100: 27.73, 26.05, 30.32, 30.58 and 29.13 dB for DiffIR, against 27.14,
25.12, 28.67, 29.83 and 28.58 dB for SRdiff [arxiv:2303.09472 tab. 2]. The
table also lists GAN restorers. It does not include ESRT or MambaIR. On real-world
×4 SR, the card records 0.74 T Mult-Adds for DiffIR against 37.25 T for LDM at
256 px input [arxiv:2303.09472 tab. 6]. The headline efficiency claim against
RePaint is shown in a figure and cannot be checked from the tables
[arxiv:2303.09472 §1]. Card:
[DiffIR: Efficient Diffusion Model for Image Restoration](/papers/2303.09472).

The released code departs from the paper's equations in several places, such as
the noise schedule, what the denoiser predicts and the channel width. An
implementation that follows only the paper will not match the released numbers.
The DiffIR card lists these differences.

### Diffusion on a fluid field (physics-informed DDPM)

This variant is not a photo model. Its DDPM is trained only on high-fidelity
two-dimensional Kolmogorov-flow vorticity, with no paired low-fidelity data
[arxiv:2211.14680 §2.2] [arxiv:2211.14680 §3.1]. At use time the coarse field is
noised to an intermediate step t, and the reverse chain runs from t to zero with
a deterministic DDIM update [arxiv:2211.14680 §2.3] [arxiv:2211.14680 alg. 2].
A sparse field is first padded by nearest-neighbour fill, and the output of each
round becomes the guide for the next. The paper uses three rounds for ×4 and
5% sparsity [arxiv:2211.14680 §2.3.1] [arxiv:2211.14680 §3.1].

The PDE residual can steer sampling in two ways. It can be fed to the noise
network as a condition, with the condition dropped at some rate during training
[arxiv:2211.14680 eq. 6] [arxiv:2211.14680 alg. 1], or it can be subtracted from
the update as a gradient step [arxiv:2211.14680 eq. 9]. The paper gives no value
for the drop rate or for the step size.

Author-reported at ×4 (64 to 256 px), the L2 norm is 0.5622 for the diffusion
model, against 0.6048 for a direct-mapping UNet and 1.3355 for bicubic. The
equation residual is 0.2178, against 0.5438 and 10.6639 [arxiv:2211.14680 tab. 1].
The L2 gap to the direct-mapping model is small, so the reported advantage rests
mainly on the equation residual and the energy spectrum [arxiv:2211.14680 §4].
At 5% sparsity the diffusion model's L2 is 0.9318 against 1.2426 for the direct
mapping [arxiv:2211.14680 tab. 1]. The authors state that one trained model
serves all these tasks without retraining [arxiv:2211.14680 §4]. Card:
[A Physics-informed Diffusion Model for High-fidelity Flow Field Reconstruction](/papers/2211.14680).

The equation residual in the paper is normalised by the vorticity norm, and the
released code does not normalise it in the same way [arxiv:2211.14680 eq. 12].
The reported equation-loss values may therefore not match a re-run, and the
card's Paper vs code section records the differences.

### Bidirectional video propagation (BasicVSR, IconVSR)

BasicVSR runs two recurrent branches over the whole clip, one forward in time and
one backward, and combines their final states for each frame [arxiv:2012.02181
§3.1] [arxiv:2012.02181 eq. 1] [arxiv:2012.02181 eq. 3]. Before a carried state
is used at the next frame, it is warped to that frame with optical flow from
SPyNet. The warp acts on features, not pixels [arxiv:2012.02181 eq. 2]
[arxiv:2012.02181 app. A]. Each branch has 30 residual blocks and 64 channels,
and the upsampler uses pixel shuffle [arxiv:2012.02181 app. A]. The authors
report that warping images instead of features costs 0.17 dB, and that removing
alignment costs 1.19 dB [arxiv:2012.02181 §3.1].

IconVSR adds two parts. At keyframes every five frames, a lightweight EDVR
extractor refills the state from three consecutive LR frames. Coupled propagation
passes the backward state into the forward branch [arxiv:2012.02181 §3.2]
[arxiv:2012.02181 eq. 4-6] [arxiv:2012.02181 app. A]. Author-reported ×4 on
REDS4 with BI degradation: BasicVSR 31.42 dB at 6.3M parameters and
63 ms, IconVSR 31.67 dB at 8.7M and 70 ms, and EDVR 31.09 dB at 20.6M and 378 ms
[arxiv:2012.02181 tab. 2]. The two IconVSR parts add 0.28 dB over BasicVSR
[arxiv:2012.02181 tab. 3]. Card:
[BasicVSR: The Search for Essential Components in Video Super-Resolution and Beyond](/papers/2012.02181).

### Second-order propagation with deformable alignment (BasicVSR++)

BasicVSR++ keeps the BasicVSR layout and changes two stages [arxiv:2104.13371
§3]. In propagation, the clip is swept four times, backward, forward, backward,
forward, so information crosses the clip twice [arxiv:2104.13371 §3.1]. Each
step uses the two most recent states of its sweep, not one, and the two are
aligned together [arxiv:2104.13371 §3.2] [arxiv:2104.13371 eq. 1]. In alignment,
the pre-trained flow sets the base of each deformable offset, and a learned
residue corrects it, so the network learns only a deviation from the flow
[arxiv:2104.13371 §3.2] [arxiv:2104.13371 eq. 6]. The authors report that
the flow base keeps training stable, where offsets learned from scratch
can overflow [arxiv:2104.13371 §1] [arxiv:2104.13371 §3.2].

The ablation adds the two parts in turn on a dataset its caption does not
name. Plain BasicVSR gives 31.48 dB, flow-guided alignment 31.94 dB, and adding
second-order propagation 32.08 dB, with the full model at 32.39 dB
[arxiv:2104.13371 tab. 3]. Without flow, alignment training collapses
at 27.44 dB [arxiv:2104.13371 tab. 4]. Author-reported ×4 on REDS4 with BI
degradation: 32.39 dB at 7.3M parameters and 77 ms, against 31.42 dB at 6.3M
and 63 ms for BasicVSR [arxiv:2104.13371 tab. 1]. The smaller BasicVSR++ (S)
reaches 32.24 dB at 6.4M parameters and 69 ms [arxiv:2104.13371 tab. 2]. The
runtime is measured at 180 × 320 LR input, and the GPU is not stated in the
table caption [arxiv:2104.13371 tab. 1]. Card:
[BasicVSR++: Improving Video Super-Resolution with Enhanced Propagation and Alignment](/papers/2104.13371).

The paper's abstract and summary headline a 0.82 dB gain over BasicVSR. That is
the (S) variant's gain, 32.24 dB against 31.42 dB, not the full model's
[arxiv:2104.13371 tab. 2]. The released code is in the MMEditing repository, now MMagic, a later snapshot
than the paper. Several details differ from the paper, and the card's Paper vs
code section lists them.

### Recurrent back-projection with optical flow (RBPN)

RBPN predicts one target frame of a video from the target and several
neighbouring frames [arxiv:1903.10128 §3]. Each neighbour is stacked with the
target and a dense optical flow between them, then passed through a projection
module that combines a single-image upscaler with a multi-frame path
[arxiv:1903.10128 §3.1] [arxiv:1903.10128 §3.2]. The module refines the
high-resolution estimate one neighbour at a time, and one convolution then
reconstructs the frame from all of the estimates [arxiv:1903.10128 §3.1]. The
authors read the module as falling back to recurrent single-image upscaling when
the neighbours add no information [arxiv:1903.10128 §3.3].

Author-reported ×4 on Vimeo-90k, PSNR: RBPN/6-PF scores 34.18, 37.28 and 40.03 dB
on slow, medium and fast clips, against 32.96, 35.84 and 37.49 dB for VSR-DUF/6
[arxiv:1903.10128 tab. 9]. On SPMCS-11 it scores 30.10 dB against 29.42 dB for
VSR-DUF [arxiv:1903.10128 tab. 8]. The authors say it beats prior methods on Vid4
except VSR-DUF [arxiv:1903.10128 §4.3]. Its Vid4 row uses a frame-dropping
protocol that differs from the usual one [arxiv:1903.10128 tab. 7]. Context
matters: on SPMCS-32, adding future frames to past ones gives 31.74 against
31.64 dB [arxiv:1903.10128 tab. 4], and removing optical flow lowers a 5-frame
model from 31.54 to 31.36 dB [arxiv:1903.10128 tab. 6]. Compute grows with the
number of neighbours; the card records 0.141 s per 4× frame for RBPN/6-PF on a
120×160 input, with the hardware not stated [arxiv:1903.10128 app. A.4]. Card:
[Recurrent Back-Projection Network for Video Super-Resolution](/papers/1903.10128).

### Anatomy-constrained SR (ACNN-SR)

The target is cardiac cine MR. Its LR stacks are 2D short-axis slices 10 mm apart,
and the HR volumes have 2 mm slices [arxiv:1705.08302 §3.1.1]. The SR network
is trained with a Huber reconstruction term, plus a shape term. A predictor maps
the output volume and the true HR volume to 64-d codes, and the squared distance
between the two codes is penalised [arxiv:1705.08302 §2.4] [arxiv:1705.08302
eq. 2]. The predictor is trained in two stages, with an autoencoder fitted to label
maps [arxiv:1705.08302 §2.4]. The predictor is not used at inference
[arxiv:1705.08302 §2.4], so the network runs as the baseline does.

Author-reported on 200 test volumes: SSIM 0.796 for ACNN-SR against 0.783 for
SR-CNN and 0.784 for 3D-UNet, and mean opinion scores of 4.36 and 4.25 from two
cardiologists against 3.59 and 3.85 for SR-CNN [arxiv:1705.08302 tab. 3]. Inference
on a 120×120×12 LR input takes 0.06 s, against 0.29 s for SR-CNN on a GTX-1080
[arxiv:1705.08302 tab. 3]. The same shape term also appears in the paper's
segmentation model, with endocardial Dice of 0.939 against 0.923 for 3D-UNet
[arxiv:1705.08302 tab. 1]. Card:
[Anatomically Constrained Neural Networks (ACNN)](/papers/1705.08302).

### Wavelet implicit network (WIRE)

WIRE is a coordinate network: an MLP maps a pixel position to an intensity. Its
activation is a complex Gabor wavelet, a complex exponential under a Gaussian
window [arxiv:2301.05187 §3.2] [arxiv:2301.05187 eq. 2]. For super-resolution,
the network is fitted by gradient descent to the low-resolution measurements,
with 4× downsampling as the forward model [arxiv:2301.05187 §4.2]. For one
single image, the paper calls its result "quantitatively higher" than the
baselines and the sharpest shown, but the text gives no PSNR number
[arxiv:2301.05187 §4.2] [arxiv:2301.05187 fig. 7]. For 4× multi-frame SR from
four Kodak frames with known sub-pixel shifts and rotations, it reports at least
1 dB PSNR and 0.04 SSIM above the other nonlinearities [arxiv:2301.05187 app. A.4].
The paper does not state w0, s0 or the learning rate for the single-image
experiment [arxiv:2301.05187 §4.2]. Card:
[WIRE: Wavelet Implicit Neural Representations](/papers/2301.05187).

### Explaining an SR network (LAM)

LAM is an attribution method for SR networks, not a new architecture. For one
output patch it gives a heat map over the LR input pixels that drive that patch
[arxiv:2011.11036 §3.3]. It is integrated gradients with two changes: the
baseline is the LR image blurred by a Gaussian, and the path progressively
sharpens the blur, instead of using a black baseline and a straight line
[arxiv:2011.11036 §3.2] [arxiv:2011.11036 §3.3]. Each map is summarised by a
diffusion index, a rescaled Gini coefficient, where a larger value means more
input pixels are involved [arxiv:2011.11036 eq. 6].

On 150 ×4 test sub-images, the authors report Pearson and Spearman correlations
above 0.8 between this index and PSNR, and between it and LPIPS, across 48
networks [arxiv:2011.11036 §4.1] [arxiv:2011.11036 §4.4]. Receptive field does
not track the index: RRDBNet's receptive field already exceeds the test-image
size, and several higher-PSNR networks do not use significantly more pixels
[arxiv:2011.11036 tab. 1] [arxiv:2011.11036 §4.4]. Card:
[Interpreting Super-Resolution Networks with Local Attribution Maps](/papers/2011.11036).

### Estimating an unknown blur kernel (KernelGAN)

KernelGAN is a front end for blind SR. It estimates the downscaling kernel of
one LR image from that image alone, with no training data, and the kernel is
then given to a non-blind SR method such as ZSSR or SRMD [arxiv:1909.06581 §2]
[arxiv:1909.06581 §5.1]. Two networks train on the input: a generator that
downscales it and a patch discriminator that scores crops of the input and of
the output [arxiv:1909.06581 §2] [arxiv:1909.06581 §3]. The generator has no
activations. It has five 64-channel hidden layers with kernel sizes 7, 5, 3, 1
and 1, and a stride-2 output layer [arxiv:1909.06581 §4.1]
[arxiv:1909.06581 fig. 4]. The product of its filters is the kernel, 13×13 at
×2 [arxiv:1909.06581 §4.2]. A regulariser keeps the kernel summing to one, with
its mass near the centre and sparse [arxiv:1909.06581 §4.2]
[arxiv:1909.06581 eq. 3]. The ×4 kernel is derived from the ×2 kernel, not
estimated again [arxiv:1909.06581 §4.2] [arxiv:1909.06581 app.].

Training runs 3,000 alternating generator and discriminator steps with Adam
(β1 0.5, β2 0.999), and a learning rate of 2e-4 divided by ten every 750 steps
[arxiv:1909.06581 §2]. The method assumes one kernel for the whole image
[arxiv:1909.06581 §5.3]. Author-reported on 100 DIV2KRK images, with ZSSR as the
SR step: 30.363 dB / 0.8669 at ×2 and 26.810 dB / 0.7316 at ×4
[arxiv:1909.06581 tab. 1]. The same ZSSR with the Michaeli and Irani kernel
scores 29.368 dB and 26.080 dB, about 1.0 dB and 0.73 dB below
[arxiv:1909.06581 tab. 1]. The ground-truth kernel with ZSSR is the upper bound
at 32.436 dB for ×2 [arxiv:1909.06581 tab. 1].

Each image is a full training run: 61 s on a V100 and 102 s on a K80, against
45 minutes per image reported for the earlier estimator [arxiv:1909.06581 §5.3].
Real photos are judged only by eye, with no numbers [arxiv:1909.06581 §5.3]
[arxiv:1909.06581 fig. 1]. The depth ablation gives 3.8 dB at ×2 and 1.6 dB at
×4 over a single-layer generator, but the single-layer model is not described,
so that gain is hard to reproduce [arxiv:1909.06581 fig. 6]. The paper's
printed eq. 2 has a sign convention that the text does not match, and the card
records the convention the authors' code uses [arxiv:1909.06581 eq. 2]. The
code is licensed for academic and non-commercial use only, so a commercial
product needs its own implementation. Card:
[Blind Super-Resolution Kernel Estimation using an Internal-GAN](/papers/1909.06581).

### Learned no-reference quality metric

Full-reference metrics need a ground-truth HR image. The authors show a
half-pixel-shifted SR image that looks fine to people but scores poorly on PSNR,
SSIM and IFC [arxiv:1612.05890 fig. 1] [arxiv:1612.05890 tab. 1]. This metric
is trained on human ratings instead. Each of 1,620 SR images has 50 absolute
ratings from 0 to 10, and its score is the mean of the median 40
[arxiv:1612.05890 §3]. The image is described by three feature groups: local
DCT statistics, global steerable-pyramid statistics and singular values of
image patches [arxiv:1612.05890 §4]. One random-forest regressor is trained per
group, and a linear layer combines the three outputs [arxiv:1612.05890 §4.4]
[arxiv:1612.05890 eq. 7].

Author-reported Spearman correlation with human scores is 0.931 overall under
5-fold cross-validation. The best baseline overall is CORNIA at 0.919, and PSNR
is at 0.604 [arxiv:1612.05890 tab. 7]. Card:
[Learning a No-Reference Quality Metric for Single-Image Super-Resolution](/papers/1612.05890).

## When to use which

| Situation | Use | Why |
|---|---|---|
| Small model (under 1M parameters), training memory is the constraint | ESRT | 751K parameters at ×4 [arxiv:2108.11084 tab. 1]; 4,191M vs 16,057M training memory against a plain Transformer baseline of unstated configuration [arxiv:2108.11084 tab. 4] |
| Lightweight SR on DIV2K only, and you can build the Mamba scan kernel | MambaIR-light | Urban100 ×4 26.52 dB vs SwinIR-light 26.47 dB, but at 42.3G vs 31.8G MACs [arxiv:2402.15648 tab. 4]. The gain is small and the cost is higher |
| Classic SR at the paper's scales, trained on DIV2K and Flickr2K with multi-GPU training | MambaIR | ×4 Set5 33.03 dB vs SwinIR 32.92 dB [arxiv:2402.15648 tab. 3]; trained on 8 V100 GPUs [arxiv:2402.15648 §4.1] |
| One backbone for several restoration tasks (denoising, deblurring, SR), and the budget allows about 20M parameters for the large model | GRL-B | Colour denoising, GoPro deblurring and ×4 Urban100 are each reported against SwinIR, Restormer or EDT [arxiv:2303.00748 tab. 2] [arxiv:2303.00748 tab. 4] [arxiv:2303.00748 tab. 5]. GRL-B has 19.81M parameters against 26.13M for Restormer in the denoising table [arxiv:2303.00748 tab. 4]. Training settings differ by task in the released code, so a re-run needs the card's Paper vs code section |
| Real photos with unknown or mixed degradations, and perceptual quality matters more than PSNR | StableSR | Best FID on DIV2K-Valid, 24.44 vs 26.47 for LDM [arxiv:2305.07015 tab. 1]. Costs 15.16 s per 512 px image at 200 steps on one V100, or 0.83 s with the 4-step SD-Turbo variant [arxiv:2305.07015 tab. 5]. Measure on your own images, since the paper's user study covers 40 internet images [arxiv:2305.07015 §4.3] |
| Fidelity to the input (PSNR, SSIM) is the target | ESRT or MambaIR, on their own benchmarks | StableSR's PSNR and SSIM are below several baselines [arxiv:2305.07015 tab. 1], and its DIV2K-Valid PSNR peaks at 24.17 dB, at w = 0.75 [arxiv:2305.07015 tab. 4]. The classic cards report PSNR on classic benchmarks, a different test set [arxiv:2108.11084 tab. 1] [arxiv:2402.15648 tab. 3] |
| Building a text-to-image pipeline with a diffusion SR stage | eDiff-I stage layout | Base at 64 px, then 256 px and 1024 px SR stages, with DEIS sampling [arxiv:2211.01324 §4.2] [arxiv:2211.01324 tab. 7]. No SR-only metric is reported, so measure your own output [arxiv:2211.01324 tab. 1] |
| Real photos with unknown degradations, scenes with many objects, a text-conditioned model is acceptable, and a 2.3B-parameter model at about 7 s per 512 px image fits the budget | SeeSR | Best FID on DIV2K-Val (31.93 vs 36.57 for StableSR) and best CLIPIQA and MUSIQ on all four sets [arxiv:2311.16518 tab. 2]. PSNR and SSIM are below the best baseline on paired sets [arxiv:2311.16518 tab. 2]. Needs a separate tagger fine-tuning stage [arxiv:2311.16518 §3.2] |
| Diffusion-quality output with few sampling steps, paired training data, and a two-stage training run is acceptable | DiffIR | Four diffusion steps on a 256-value prior [arxiv:2303.09472 §5.1]. Real-world ×4 at 0.74 T Mult-Adds against 37.25 T for LDM [arxiv:2303.09472 tab. 6]. Its SR table has no ESRT or MambaIR baseline, so its PSNR is not ranked against them [arxiv:2303.09472 tab. 2] |
| Simulated or measured 2D fields with a known governing equation, high-fidelity training fields, and no paired coarse data | Physics-informed DDPM (flow-field card) | 4× L2 0.5622 vs 0.6048 for a direct-mapping UNet, and equation residual 0.2178 vs 0.5438 [arxiv:2211.14680 tab. 1]. Tested on one flow family at Re 1000, so the gain on other flows is not shown. The equation metric differs from the code's [arxiv:2211.14680 eq. 12] |
| Video, neighbouring frames available, optical flow can be estimated for each neighbour, and the benchmark is Vimeo-90k or SPMCS-11 | RBPN | Vimeo-90k fast clips 40.03 dB against 37.49 dB for VSR-DUF/6 [arxiv:1903.10128 tab. 9]; SPMCS-11 30.10 dB against 29.42 dB [arxiv:1903.10128 tab. 8]. Cost rises with neighbour count [arxiv:1903.10128 app. A.4]. No REDS4 result is reported, so it cannot be ranked against BasicVSR++ on REDS4 |
| Video with many frames of one scene, and the whole clip is available before output, and the best REDS4 score matters more than the simplest model | BasicVSR++ | REDS4 BI: 32.39 dB at 77 ms and 7.3M parameters, against IconVSR 31.67 dB at 70 ms and 8.7M [arxiv:2104.13371 tab. 1] [arxiv:2012.02181 tab. 2]. Needs flow and deformable convolution, so it is more work to implement [arxiv:2104.13371 §3.2] |
| Video with many frames of one scene, and the whole clip is available before output, and a simpler design is preferred | BasicVSR, or IconVSR for a small gain | REDS4 BI: BasicVSR 63 ms at 6.3M parameters vs EDVR 378 ms at 20.6M; IconVSR adds 0.25 dB at 70 ms and 8.7M [arxiv:2012.02181 tab. 2]. Output frames need the whole clip, since both sweeps run over all frames [arxiv:2012.02181 eq. 1] [arxiv:2012.02181 eq. 3]. Memory for long videos is not reported [arxiv:2012.02181 app. A] |
| Volumetric medical SR where anatomical plausibility matters and inference cost must stay the same | ACNN-SR | The shape term is dropped at inference [arxiv:1705.08302 §2.4]; SSIM 0.796 vs 0.783 for SR-CNN, and 0.06 s vs 0.29 s per 120×120×12 input [arxiv:1705.08302 tab. 3]. The SR experiment covers cardiac MR only [arxiv:1705.08302 §3.5] |
| Few LR frames of one scene with known sub-pixel shifts, and per-signal fitting is acceptable | WIRE | Multi-frame 4× SR at least 1 dB above other INR activations [arxiv:2301.05187 app. A.4]. The SR experiments use one image or one frame set each, so generalisation is not shown [arxiv:2301.05187 §4.2]. No SR run time is reported; the Kodak image fit takes 1.6 min [arxiv:2301.05187 §4.1] |
| Need to know which input pixels an SR network uses for one output patch | LAM | Gradient-based attribution with a blurred baseline [arxiv:2011.11036 §3.3]. It needs one forward and backward pass of the full image per path step, with 100 steps used in the paper [arxiv:2011.11036 §3.3] |
| Blur of the LR image is unknown, the input is one image, and a per-image run of about a minute is acceptable | KernelGAN, then ZSSR or SRMD with the estimated kernel | 30.363 dB / 0.8669 at ×2 with ZSSR, against 29.368 dB for the Michaeli and Irani kernel [arxiv:1909.06581 tab. 1]. About 61 s per image on a V100 [arxiv:1909.06581 §5.3]. Assumes one kernel for the whole image [arxiv:1909.06581 §5.3]. Non-commercial code, so a product needs its own implementation (see the card) |
| No sharp reference image, and the target is what people rate highly | The no-reference metric | Spearman 0.931 with human scores under 5-fold CV [arxiv:1612.05890 tab. 7]; 0.852 with 6 reference images held out [arxiv:1612.05890 tab. 8]. Costs 13.31 s per image in MATLAB [arxiv:1612.05890 §5.3], and assumes the blur kernel width is known for each scale [arxiv:1612.05890 §3] |
| Ground-truth HR image available, and you need a fast score | PSNR or SSIM | Timing of 0.02 s and 0.13 s per image [arxiv:1612.05890 §5.3], against Spearman with human scores of 0.604 and 0.635 overall [arxiv:1612.05890 tab. 7] |

The classic reconstruction cards do not compare with each other, so their
Urban100 numbers are not a head-to-head ranking. ESRT, MambaIR-light and GRL-B
report ×4 Urban100 scores of 26.39, 26.52 and 28.53 dB, each in its own table
against its own baselines [arxiv:2108.11084 tab. 1] [arxiv:2402.15648 tab. 4]
[arxiv:2303.00748 tab. 5]. The StableSR, BasicVSR, ACNN, DiffIR and RBPN cards each use their own baselines
and metrics, so their numbers are not placed on one scale either.

## Implementation sketch

The classic reconstruction variants share this training core. Only the body
changes. ACNN-SR keeps the Huber reconstruction term and adds a training-only
shape term. The diffusion stages, BasicVSR, WIRE, LAM and the no-reference
metric do not use this loop.

```pseudocode
# Common SR training step  [arxiv:2108.11084 §4.2] [arxiv:2402.15648 §4.1]
# Input:  LR/HR patch pair cut from DIV2K; LR patch is 48x48 in ESRT
# Output: updated model parameters
def train_step(model, lr_patch, hr_patch, optimizer):
    # feature extraction, body, then upsampling by the scale factor
    # (the body is the part that differs between variants)
    sr = model(lr_patch)                    # (B, 3, s*h, s*w)
    # L1 loss against the ground-truth HR patch  [arxiv:2108.11084 §4.2]
    loss = mean(abs(sr - hr_patch))
    optimizer.zero_grad()
    loss.backward()
    # Adam, initial learning rate 2e-4  [arxiv:2108.11084 §4.2] [arxiv:2402.15648 §4.1]
    optimizer.step()
    return loss
```

Where the variants differ in the body:

- **ESRT:** convolutional High Preserving Blocks, then one Efficient
  Transformer over unfolded 3×3 tokens, with attention split into four segments
  [arxiv:2108.11084 §3.3.2].
- **MambaIR:** residual state-space blocks. Each runs a selective scan over four
  directions of the feature map, then a convolution bottleneck with channel
  attention [arxiv:2402.15648 §3.3] [arxiv:2402.15648 §3.5].
- **GRL (this loop, with the paper's stated values):** the paper gives L1 loss,
  Adam and an initial learning rate of 2e-4 [arxiv:2303.00748 §5]. The
  schedule and the batch and patch sizes are not stated in the paper. Its body
  is the three-branch layer described above, with window, stripe and
  convolution parts run in parallel [arxiv:2303.00748 §4].
- **Physics-informed DDPM (not this loop):** the noise network is trained with a
  squared-error loss on the noise, as in the diffusion cards, and with a drop
  of the PDE condition [arxiv:2211.14680 alg. 1]. Sampling starts from the noised
  guide at step t, not from pure noise, and runs DDIM with sigma zero
  [arxiv:2211.14680 alg. 2]. The paper gives the step counts for each task, but
  not the drop rate or the guidance strength [arxiv:2211.14680 §3.1].
- **BasicVSR++ (video, not this loop):** a Charbonnier loss, with Adam on
  30-frame REDS samples and batch size 8 for 600K iterations
  [arxiv:2104.13371 app. B] [arxiv:2104.13371 §4]. The main network uses a
  learning rate of 1e-4, the flow network 2.5e-5, and the flow network is fixed
  for the first 5,000 iterations on REDS [arxiv:2104.13371 §4]
  [arxiv:2104.13371 app. B]. The paper does not state the Charbonnier epsilon
  or the Adam betas [arxiv:2104.13371 app. B].
- **ACNN-SR (this loop plus a shape term):** the Huber reconstruction term, plus
  λ1 times the squared distance between predictor codes of the output and the
  true HR volume, plus weight decay. The paper gives λ1 = 0.01 and weight decay
  5e-6, and a fixed learning rate of 0.001 with the optimiser variant not stated
  [arxiv:1705.08302 eq. 2] [arxiv:1705.08302 supp. Implementation Details].
- **BasicVSR (video, not this loop):** a Charbonnier loss with ε = 1e-8,
  trained with Adam on 15-frame REDS clips and 14-frame Vimeo-90K
  clips, batch size 8, 300K iterations [arxiv:2012.02181 app. A] [arxiv:2012.02181
  eq. 7] [arxiv:2012.02181 §4]. Initial learning rates are 2e-4 for most modules,
  1e-4 for the feature extractor and 2.5e-5 for the flow estimator
  [arxiv:2012.02181 §4].
- **StableSR (diffusion, not this loop):** the loss is the mean squared error
  between the predicted and the true noise on the latent, with the U-Net frozen.
  Only the time-aware encoder and the modulation layers are updated, with Adam at
  learning rate 5e-5 [arxiv:2305.07015 §3.1] [arxiv:2305.07015 §4.1]
  [arxiv:2305.07015 §4.3]. At inference, the noise predictions for overlapping
  64×64 latent patches are blended with Gaussian weights at every step
  [arxiv:2305.07015 alg. 1] [arxiv:2305.07015 eq. 7].
- **SeeSR (diffusion with tag and embedding prompts, not this loop):** the tagger
  is first trained with an MSE term on representation embeddings and a
  cross-entropy term on the tag logits, both against the frozen tagger's output
  on the clean image [arxiv:2311.16518 eq. 1] [arxiv:2311.16518 §3.2]. The
  diffusion stage then minimises the noise-prediction MSE [arxiv:2311.16518 eq. 2],
  with the UNet, text encoder and VAE frozen. Only the image encoder, ControlNet
  and the cross-attention modules are updated [arxiv:2311.16518 §3.3]. The paper reports Adam at 5e-5 for the diffusion
  stage, and 50 spaced DDPM steps at inference [arxiv:2311.16518 §4.1].
- **WIRE (not this loop):** the fit is an l2 loss between the forward-modelled
  output and the measurements, with Adam and no other regularisation, and the
  learning rate decays to 0.1× its start [arxiv:2301.05187 §4]. There is no
  ground-truth HR patch in the loss.
- **KernelGAN (kernel estimation, not this loop):** the generator and
  discriminator are updated in turn for 3,000 steps, with Adam and the learning
  rate schedule given under Variants [arxiv:1909.06581 §2]. The generator's loss
  adds a regulariser on the kernel [arxiv:1909.06581 eq. 3]. The output is one
  kernel, not an SR image, so the SR step runs afterwards with that kernel.
- **DiffIR (diffusion on a prior, not this loop):** stage 1 trains the prior
  encoder and the transformer with an L1 loss on the restored image. Stage 2
  adds an L1 loss on the prior and trains the denoiser on noised prior vectors
  [arxiv:2303.09472 §4.1] [arxiv:2303.09472 §4.2] [arxiv:2303.09472 eq. 13]. The
  paper gives Adam with betas 0.9 and 0.99, learning rate 2e-4 and batch 64 for
  SR, for 1M iterations [arxiv:2303.09472 §5.1] [arxiv:2303.09472 app. E]. It
  states no learning-rate schedule for SR [arxiv:2303.09472 §5]. At inference,
  the prior is sampled from a standard normal and denoised over the four steps,
  then passed to the transformer [arxiv:2303.09472 alg. 2].
- **RBPN (video, not this loop):** a per-pixel L1 loss with Adam (momentum 0.9)
  and an initial learning rate of 1e-4, batch size 8, 150 epochs on Vimeo-90k
  7-frame sequences [arxiv:1903.10128 §4.1]. The paper's wording on when the
  rate drops is ambiguous, so the RBPN card records how the released code does
  it.
- **Test-time memory:** ESRT's memory figures come from 48×48 training patches.
  The paper does not give memory for full-size inputs [arxiv:2108.11084 tab. 4],
  so plan tiling or chopping at inference time and measure it yourself.
- **Evaluation, not training:** the no-reference metric is a separate pass over
  a finished SR output. Its feature extraction and forest prediction are
  sketched on its card [arxiv:1612.05890 alg. 1], and the card's known gaps are
  listed under its Paper vs code section.

## Contradictions

- **SwinIR scale.** ESRT lists SwinIR under ×4 with BSD100 29.20 dB and
  Manga109 33.98 dB [arxiv:2108.11084 tab. 7]. MambaIR lists the same two values
  for SwinIR-light at ×3, and gives different ×4 values for SwinIR-light
  [arxiv:2402.15648 tab. 4]. The two papers use different SwinIR variants, and the
  scale label in ESRT's table cannot be settled from these two cards.
- **ESRT's own numbers.** ESRT reports 770K parameters at ×3 and 751K at ×4
  [arxiv:2108.11084 tab. 1], and 0.68M in its timing table [arxiv:2108.11084 tab. 2].
  Its ×3 row (770K, BSD100 29.15 dB) reappears as the ESRT row of the table labelled
  ×4 [arxiv:2108.11084 tab. 7]. The paper does not resolve these.
- **Agreement on shared baselines.** The baselines the two classic reconstruction
  cards share match. IMDN at ×4 on Urban100 is 26.04 dB in both, and LatticeNet
  is 26.25 dB in both [arxiv:2108.11084 tab. 1] [arxiv:2402.15648 tab. 4].
  GRL's table has the same SwinIR rows: 27.45 dB at Urban100 ×4 for the larger
  model and 26.47 dB for the 0.9M model [arxiv:2303.00748 tab. 5], matching
  MambaIR's SwinIR figures [arxiv:2402.15648 tab. 3] [arxiv:2402.15648 tab. 4].
- **Diffusion stages and reconstruction models.** eDiff-I's SR stages are not
  compared with ESRT or MambaIR, and its paper gives no SR-only metric
  [arxiv:2211.01324 tab. 1]. StableSR's baselines are RealSR, BSRGAN, DASR, Real-ESRGAN+,
  FeMaSR, LDM, SwinIR-GAN and IF_III, not ESRT or MambaIR [arxiv:2305.07015 tab. 1].
  SeeSR's baselines are BSRGAN, Real-ESRGAN, LDL, DASR, FeMaSR, LDM, StableSR,
  ResShift, PASD and DiffBIR, and they also exclude ESRT and MambaIR
  [arxiv:2311.16518 §4.1]. No head-to-head number exists across these cards.
- **StableSR's FID on DIV2K-Val.** The StableSR paper reports FID 24.44 for
  StableSR against 26.47 for LDM on DIV2K-Valid [arxiv:2305.07015 tab. 1]. The
  SeeSR paper reports 36.57 for StableSR on DIV2K-Val, the best other method
  there, against 31.93 for SeeSR [arxiv:2311.16518 tab. 2]. The two cards do
  not show that the test crops and degradations match, so the two StableSR
  values are not on one scale.
- **Protocols, not one scale.** StableSR's PSNR is measured on DIV2K-Valid with
  Real-ESRGAN degradations and on real-world sets [arxiv:2305.07015 §4.2]. The
  classic cards use classic benchmarks [arxiv:2108.11084 tab. 1]. KernelGAN's
  DIV2KRK uses synthetic anisotropic Gaussian kernels [arxiv:1909.06581 §5.2].
  Their PSNR values are not comparable, and neither are the ACNN opinion scores.
  The flow-field card's L2 and equation residual are measured on vorticity
  fields, so they are not comparable with any PSNR on this page
  [arxiv:2211.14680 tab. 1].
- **Kernel estimation and SR scores.** KernelGAN's ×2 and ×4 numbers depend on
  which non-blind SR method runs after it: with SRMD, KernelGAN scores 29.565 dB
  at ×2, and with ZSSR 30.363 dB [arxiv:1909.06581 tab. 1]. Compare kernels only
  with the same SR step. Its ×4 SRMD result is below PDN in PSNR, but above it in
  SSIM [arxiv:1909.06581 tab. 1].
- **BasicVSR: text against table.** The text says BasicVSR beats prior methods on
  REDS4, UDM10 and Vid4 [arxiv:2012.02181 §4]. In the table, EDVR scores higher on
  Vid4 BI (27.35 dB against 27.24 dB) and on Vimeo-90K-T BI (37.61 dB against 37.18
  dB) [arxiv:2012.02181 tab. 2]. Quote the table column, not the summary.
- **ACNN speed claim.** The text says ACNN-SR is faster "by a factor of 5"
  [arxiv:1705.08302 §3.5]. Its table gives 0.29 s against 0.06 s, which is about
  4.8× [arxiv:1705.08302 tab. 3].
- **RBPN flow timing.** The paper says the optical flow is precomputed
  [arxiv:1903.10128 §3.1]. The RBPN card records that the released code computes
  it for each training sample during loading. The two differ in training cost and
  in the augmentation the flow sees.
- **DiffIR text and table.** The SR section lists Set5 as a test set, but the SR
  table has no Set5 column [arxiv:2303.09472 §5.3] [arxiv:2303.09472 tab. 2]. The
  table caption names LPIPS and DISTS, while its columns are PSNR and LPIPS
  [arxiv:2303.09472 tab. 2]. Quote the columns.
- **DiffIR and the classic cards.** DiffIR's Urban100 ×4 score of 26.05 dB is
  below ESRT's 26.39 dB and MambaIR-light's 26.52 dB [arxiv:2108.11084 tab. 1]
  [arxiv:2402.15648 tab. 4] [arxiv:2303.09472 tab. 2]. The three papers use
  different baselines and protocols, so the gap does not rank them.
- **WIRE and the reconstruction models.** WIRE's SR is compared only with other
  INR activations (SIREN, Gaussian, ReLU with positional encoding, MFN), not with
  ESRT or MambaIR [arxiv:2301.05187 §4] [arxiv:2301.05187 app. A.4]. Its
  single-image SR gives no number in the text, so it cannot be placed on the
  benchmark scale used by the two classic reconstruction cards [arxiv:2301.05187 §4.2].

## Papers

- [Efficient and Explicit Modelling of Image Hierarchies for Image Restoration](/papers/2303.00748): a restoration transformer whose layers run window, anchored stripe and convolution branches in parallel.
- [A Physics-informed Diffusion Model for High-fidelity Flow Field Reconstruction](/papers/2211.14680): a DDPM trained on high-fidelity vorticity that reconstructs coarse or sparse flow fields from a noised guide, optionally steered by the PDE residual.
- [Recurrent Back-Projection Network for Video Super-Resolution](/papers/1903.10128): a recurrent video SR network that refines high-resolution features one optical-flow-aligned neighbour at a time.
- [DiffIR: Efficient Diffusion Model for Image Restoration](/papers/2303.09472): a diffusion model that denoises a compact 256-value restoration prior in four steps, then a transformer restores the image.
- [Blind Super-Resolution Kernel Estimation using an Internal-GAN](/papers/1909.06581): a per-image linear-generator GAN that estimates the downscaling kernel, which a non-blind SR method then uses.
- [Learning a No-Reference Quality Metric for Single-Image Super-Resolution](/papers/1612.05890): a learned, reference-free score trained on human ratings, from three random-forest feature groups.
- [Interpreting Super-Resolution Networks with Local Attribution Maps](/papers/2011.11036): an integrated-gradients attribution for SR networks, with a blurred baseline and progressive path.
- [Transformer for Single Image Super-Resolution](/papers/2108.11084): ESRT, a lightweight CNN-Transformer hybrid with a memory-reduced attention.
- [eDiff-I: Text-to-Image Diffusion Models with an Ensemble of Expert Denoisers](/papers/2211.01324): a text-to-image cascade whose 256 px and 1024 px stages are diffusion SR models.
- [SeeSR: Towards Semantics-Aware Real-World Image Super-Resolution](/papers/2311.16518): Stable Diffusion SR steered by tag text and tagger embeddings from a degradation-aware tagger, with an LR-noised start.
- [MambaIR: A Simple Baseline for Image Restoration with State-Space Model](/papers/2402.15648): restoration backbone that replaces window attention with a four-direction state-space scan.
- [WIRE: Wavelet Implicit Neural Representations](/papers/2301.05187): an MLP whose Gabor-wavelet activation fits images, including SR, from their measurements.
- [Exploiting Diffusion Prior for Real-World Image Super-Resolution](/papers/2305.07015): StableSR, a time-aware encoder that steers a frozen Stable Diffusion for real photos, with patch-blended sampling.
- [BasicVSR: The Search for Essential Components in Video Super-Resolution and Beyond](/papers/2012.02181): bidirectional recurrent video SR with flow-warped features, plus IconVSR's keyframe refill.
- [Anatomically Constrained Neural Networks (ACNN)](/papers/1705.08302): a training-only code-distance shape term for cardiac MR super-resolution and segmentation.
- [BasicVSR++: Improving Video Super-Resolution with Enhanced Propagation and Alignment](/papers/2104.13371): second-order grid propagation with flow-guided deformable alignment, extending BasicVSR.