AVSR-Diff: Scale-Agnostic Diffusion Priors for Temporally Consistent Arbitrary-Scale Video Super-Resolution

Conference: ECCV 2026
arXiv: 2607.00987
Code: None (Project page: https://kaist-viclab.github.io/AVSR-Diff/)
Area: Diffusion Models / Video Super-Resolution
Keywords: Video Super-Resolution, Diffusion Models, Arbitrary-Scale, Temporal Consistency, Implicit Neural Representation

TL;DR

This paper proposes AVSR-Diff, which decouples the scale-agnostic latent denoising of diffusion models from continuous coordinate decoding. It utilizes a Temporal Gated Feature Recurrence (TGFR) module for cross-frame feature alignment to suppress flickering, and a Scale-Aware Fourier Refinement (SAFR) module for scale-adaptive frequency modulation in the Fourier domain. This architecture enables the simultaneous achievement of high generation quality and stable temporal consistency for arbitrary-scale VSR, while maintaining a computational overhead that does not scale with the output resolution.

Background & Motivation

Diffusion model-driven video super-resolution (DM-based VSR) has achieved substantial improvements in perceptual quality at fixed integer upscaling factors (typically 4x). However, real-world deployment frequently demands continuous and arbitrary resolution scaling—and training a separate diffusion model for each target scale is highly impractical. On the other hand, coordinate-based Implicit Neural Representation (INR) methods naturally support continuous querying at arbitrary scales, but they are limited by L1/L2 regression losses. Consequently, they lose high-frequency details at larger scaling factors (e.g., 6x, 8x), yielding over-smoothed video frames.

A natural idea is to combine the generative priors of diffusion models with the continuous decoding capability of INRs. While this path has been explored in single-image super-resolution, extending it to video encounters a non-trivial bottleneck: the frame-wise randomness inherent in diffusion sampling causes subtle drifts in inter-frame features. A continuous coordinate decoder is highly sensitive to such feature instability, and even minor drifts are amplified into severe temporal flickering. Although existing DM-based VSR methods incorporate mechanisms like warping guidance or temporal-attention blocks to mitigate flickering, they are designed for fixed scales and cannot provide the scale-agnostic, strictly aligned latent features required for continuous decoding. Furthermore, directly applying a full video diffusion model (such as VEnhancer) to perform 3D U-Net denoising at the target resolution results in memory and computational overheads that explode with larger upscaling factors, making large-scale inference highly impractical.

Core Idea: Completely decouple "generative prior extraction" from "resolution rendering." Diffusion sampling is performed exclusively in a fixed, low-resolution latent space, ensuring that the generation cost remains independent of the target scale. Subsequently, a carefully designed continuous video decoder renders pixels at arbitrary resolutions on top of these scale-agnostic latent features, while strictly suppressing temporal flickering through gated feature recurrence and Fourier-domain frequency modulation.

Method

Overall Architecture

AVSR-Diff is built upon a pre-trained SD×4 Upscaler (Stable Diffusion 4x super-resolution model), freezing its VAE encoder-decoder pair \((E, D)\) and denoising U-Net \((\epsilon_\theta)\), while only training an additionally introduced ControlNet \((C_\phi)\) and a continuous video decoder \((D_s)\). The entire pipeline is divided into two strictly decoupled stages:

First Stage: Scale-Agnostic Latent Denoising. Given a low-resolution (LR) video sequence \(x = \{x^i\}\), at each diffusion denoising step, the LR frame \(x^i\) is directly concatenated along the channel dimension with the noisy latent \(z_t^i\) and fed into both the ControlNet and the frozen U-Net. The TGFR module inside the ControlNet recurrently propagates and aligns deep residual features across adjacent frames, supplying the U-Net with strictly aligned temporal conditions. This process eventually yields the denoised latent sequence \(z_0 = \{z_0^i\}\). Throughout this stage, the spatial resolution remains strictly in the LR latent space (e.g., \(64 \times 64\)), completely independent of the target upscaling factor.

Second Stage: Continuous Arbitrary-Scale Decoding. The denoised latents \(z_0\) are fed into the extended continuous video decoder \(D_s\). First, intermediate deep features \(F^i\) are extracted via the frozen VAE decoder and modulated through scale-aware Group Normalization (GN). Next, the TGFR module is leveraged again for bidirectional feature propagation to reinforce temporal consistency. Subsequently, the SAFR module dynamically modulates the spectral components in the Fourier domain according to the target scale. Finally, pixels of any scale are rendered using LIIF-style coordinate queries coupled with bilinear residual connections.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["LR Video Sequence x"] --> B["Scale-Agnostic Latent Denoising<br/>Frozen U-Net + Trainable ControlNet"]
    B --> C["TGFR: Temporal Gated Feature Recurrence<br/>DCN Sub-pixel Alignment + Gated Fusion"]
    C --> D["Denoised Latent Sequence z0"]
    D --> E["Continuous Video Decoder<br/>Scale-Aware Feature Adaptation + TGFR Bidirectional Propagation"]
    E --> F["SAFR + LIIF Rendering<br/>Fourier-Domain Spectral Gating -> Continuous Coordinate Queries"]
    F --> G["Arbitrary-Scale HR Video"]

Key Designs

1. TGFR: Temporal Gated Feature Recurrence—Mitigating Inter-Frame Flickering Caused by Diffusion Sampling

In frame-by-frame sampling of diffusion models, even under identical LR conditions, random noise injected at each step introduces subtle variations in the latent features of adjacent frames. Continuous coordinate decoders are highly sensitive to these discrepancies—a pixel-level drift in the feature space can translate into severe jitter across entire texture regions in the rendered video. The objective of TGFR is to minimize this drift within the residual features of the ControlNet.

The specific implementation consists of three steps. First, coarse alignment: the clean latent estimation of the previous frame (\(i-1\)) at the current denoising step, denoted as \(\hat{z}_0^{i-1}\), is mapped back to the RGB space via the frozen decoder \(D\) to obtain \(\hat{y}_{\text{base}}^{i-1}\). This is then warped to the current frame using the optical flow \(f^{i \to i-1}\) estimated by a pre-trained RAFT, serving as an explicit pixel-level condition for the ControlNet. Second, fine alignment: the intermediate residual features \(H^{i-1}\) of the ControlNet from the previous frame are warped using the same optical flow, passed through a ResBlock, and fed into a Deformable Convolutional Network (DCN). By learning dynamic offsets \(\Delta p\) and modulation masks \(\Delta m\), it refines optical flow inaccuracies at the sub-pixel feature level. Third, gated fusion: the aligned features \(H_{\text{aligned}}^{i-1 \to i}\) are fused with the current frame's ControlNet residual features \(C^i\) in a spatially adaptive manner. The gating map is computed as \(G^i = \text{sigmoid}(\text{Conv}([C^i, H_{\text{aligned}}, |C^i - H_{\text{aligned}}|]))\), and the final output is formulated as \(H^i = C^i + G^i * H_{\text{aligned}}^{i-1 \to i}\). Crucially, the input to the gating map includes the absolute difference \(|C^i - H_{\text{aligned}}|\) between the current and aligned features. This allows the network to explicitly assess whether the features propagated from the previous frame are spatially reliable: regions with substantial discrepancies produce gate values approaching 0, automatically blocking unreliable propagated information and preventing error accumulation.

This TGFR mechanism is deployed in two places: the denoising stage (unidirectional recurrence within the ControlNet) and the decoding stage (bidirectional cascaded propagation within the continuous video decoder). It serves as the "temporal backbone" of the entire framework.

2. SAFR: Scale-Aware Fourier Refinement—Enabling scale-dependent high-frequency synthesis

Different upscaling factors demand distinct levels of high-frequency detail: a 2x scale only requires light edge sharpening to avoid over-enhancement artifacts like ringing, whereas an 8x scale demands synthesis of complex textures from near-absent information. If a continuous decoder outputs a static, scale-agnostic feature map, it inevitably leads to over-enhancement at small scales or over-smoothing at large scales. The design motivation of SAFR is to explicitly model this "scale-frequency" dependency directly in the Fourier domain.

SAFR operates on the temporally-aligned deep features \(\tilde{u}^i\). First, a 2D Fast Fourier Transform (FFT) computes the spectrum \(U^i = \mathcal{F}(\tilde{u}^i)\). A \(1 \times 1\) convolution \(\text{Mix}_C\) mixes the spectral information across channels in the complex domain (performing frequency-domain channel interaction). This mixed spectrum is then multiplied by a channel-wise spectral gating vector \(\psi(s)\), which is predicted by an MLP based on the target scale \(s\). Finally, an Inverse FFT (IFFT) projects the result back to the spatial domain, followed by a residual summation with the original features: \(u_{\text{ref}}^i = \tilde{u}^i + \text{Conv}(\mathcal{F}^{-1}(\text{Mix}_C(U^i) * \psi(s)))\). At larger scales, the high-frequency channel weights in \(\psi(s)\) are automatically amplified, intensifying the high-frequency energy in the spectrum to decode richer textures. Conversely, at smaller scales, high frequencies are suppressed to prevent over-sharpening.

In addition to SAFR, the continuous decoder implements channel-wise scale modulation at the feature extraction stage: \(u^i = \text{GN}(F^i) * (1 + \gamma(s)) + \beta(s)\), where \(\gamma(s)\) and \(\beta(s)\) are mapped from the positional encoding of scale \(s\) via an MLP. Under this dual-level scale injection (spatial-domain input modulation and frequency-domain spectral gating), the decoder's adaptation to arbitrary continuous scales is systematically constructed from coarse to fine.

3. Gated Sparsity Regularization: Ensuring Stability in Long-Sequence Inference

Recurrent feature propagation in TGFR poses a danger for long videos (e.g., 100 frames): if any frame's propagated feature contains errors, they will accumulate sequentially along the temporal axis, leading to catastrophic visual collapse. The proposed solution is to introduce a gated sparsity penalty in the ControlNet training loss: \(\mathcal{L}_{\text{gate}} = \frac{\lambda_{\text{gate}}}{|G|} \sum_{g \in G} \|g\|_1\), which applies an \(L1\) regularization to all gating maps \(G\) produced by the various scales of TGFR. This forces the network to open the gating channels only when the propagation gains clearly outweigh the activation cost, thereby suppressing redundant feature propagation and cutting off error accumulation at its source.

Ablation studies clearly demonstrate the necessity of this design: in a single forward pass over a 100-frame sequence, replacing the gate with simple concatenation (w/o Gate Concat.) causes the temporal optical flow metric (tOF) to skyrocket from 16.79 to 1887.47 (complete visual collapse). Retaining the gate but omitting the sparsity penalty (Gate w/o Sparsity) still leads to degradation after approximately 50 frames. In contrast, the full TGFR remains stable throughout the entire sequence, achieving optimal scores across all metrics.

Detailed Inference Example: 8x Super-Resolution

Taking an 8-frame \(64 \times 64\) LR video upscaled to 8x as an example, the full inference pipeline proceeds as follows. First, forward and backward optical flows are estimated between all LR frame pairs using RAFT. Second, the latent sequence is initialized as \(z_T \sim \mathcal{N}(0, I)\), and a 50-step DDPM sampling process is initiated. At each denoising step, the propagation direction alternates (forward or backward). The ControlNet receives the warped RGB anchors and residual features from the previous frame. Following DCN alignment and gated fusion, these features are injected into the frozen U-Net. The U-Net predicts the noise for the current frame, and DDPM updates the latent to \(z_{t-1}\). Third, the clean latent sequence \(z_0\) is resolved. Fourth, \(z_0\) is fed into the VAE decoder to extract intermediate features \(F^i\), which are modulated using scale-aware GN with \(s=4.0\) (clipped to the upper training boundary to avoid out-of-distribution positional encoding extrapolation artifacts). Fifth, forward and backward TGFR propagations are executed to yield temporally enriched features \(\tilde{u}^i\). Sixth, SAFR performs spectral gating modulation in the Fourier domain using \(s=4.0\). Seventh, LIIF queries local features inside \(u_{\text{ref}}^i\) coordinate-by-coordinate for the target \(512 \times 512\) resolution to predict RGB residuals, which are then added to the bilinearly upsampled LR inputs to obtain the final 8x HR frames. Crucially, the entire denoising process is performed in the \(64 \times 64\) latent space, maintaining a constant peak memory usage of approximately 8.7GB, almost identical to that of 2x upscaling.

Loss & Training

ControlNet \(C_\phi\) and the continuous video decoder \(D_s\) are trained independently using the Adam optimizer with a batch size of 32 and sequences of 8 \(64 \times 64\) LR frames.

ControlNet Training (30K steps, lr=1e-5): \(\mathcal{L}_{\text{CNet}} = \|\epsilon - \hat{\epsilon}\|_2^2 + \lambda_{\text{gate}} \cdot \text{mean}(\|G\|_1)\), where \(\epsilon\) is the ground-truth noise, \(\hat{\epsilon}\) is the noise predicted by the U-Net (conditioned on the ControlNet), and \(\lambda_{\text{gate}} = 0.01\). Inference utilizes 50-step DDPM sampling.

Continuous Decoder Training (Two-stage, 140K steps total): In the first stage (100K steps, lr=1e-4 with cosine annealing), the model is trained with only L1 loss and perceptual loss (\(\lambda_{\text{percep}} = 1.0\)). In the second stage (40K steps, lr=1e-5), a PatchGAN adversarial loss (\(\lambda_{\text{adv}} = 0.05\)) is introduced to enhance high-frequency realism. During training, the target scale \(s\) is uniformly sampled from \([1.1, 4.0]\). For inference at larger scales, a clipping strategy of \(s=4\) is applied to the scale-aware modules to avoid out-of-distribution issues.

Key Experimental Results

Main Results

The following two tables present main results comparing fixed 4x upscaling and multi-scale continuous upscaling. All methods are evaluated on the REDS4 and Vid4 datasets, utilizing perceptual metrics (LPIPS, DISTS), pixel-fidelity metrics (PSNR, SSIM), and temporal consistency metrics (tLPIPS, tOF).

Fixed 4x Scale Comparison (Table 1, REDS4):

Method	Type	LPIPS↓	DISTS↓	PSNR↑	SSIM↑	tLPIPS↓	tOF↓
BasicVSR++	Regression / Fixed	13.49	6.99	32.32	0.9057	9.19	18.16
RVRT	Regression / Fixed	13.32	6.91	32.70	0.9106	8.98	18.08
StableVSR	Generative / Fixed	9.74	4.51	27.97	0.7951	5.40	17.20
MGLD-VSR	Generative / Fixed	14.53	6.23	26.25	0.7408	16.36	39.62
STAR	Generative / Fixed	29.48	12.17	23.08	0.6726	32.98	64.53
VEnhancer	Generative / Arbitrary	34.69	14.91	22.90	0.6413	24.95	95.51
AVSR-Diff	Generative / Arbitrary	9.54	4.42	28.75	0.8204	4.20	16.79

AVSR-Diff achieves state-of-the-art performance in five out of six metrics across perceptual quality (LPIPS, DISTS) and temporal consistency (tLPIPS, tOF), while also demonstrating the highest PSNR/SSIM among generative methods. Despite being an arbitrary-scale model, it outperforms all fixed-scale generative methods specifically designed for 4x upscaling on the fixed 4x task.

Multi-Scale Arbitrary Upscaling Comparison (Table 2, REDS4, Excerpt):

Method	2x LPIPS↓	2x PSNR↑	2x tOF↓	3.25x LPIPS↓	3.25x PSNR↑	8x LPIPS↓	8x PSNR↑	8x tOF↓
SAVSR (Regression)	6.66	35.25	7.82	19.51	27.13	43.02	25.50	46.12
V3VSR (Regression)	6.01	36.13	7.95	18.17	26.23	44.25	25.89	44.51
StableVSR (Generative / Fixed+Bicubic)	5.51	33.49	8.01	14.38	24.91	34.96	23.72	44.09
VEnhancer (Generative / Arbitrary)	23.39	23.27	77.78	30.71	22.53	43.87	21.73	129.42
AVSR-Diff	3.84	35.47	7.37	8.17	26.50	29.43	24.95	39.13

Across all scales, AVSR-Diff consistently leads in perceptual quality and temporal consistency. For 2x upscaling, its PSNR (35.47) approaches that of strong regression-based baselines. At the larger 8x scale, its tOF advantage over VEnhancer is as high as 3.3x.

Ablation Study

Core Component Ablation (Table 3, REDS4 4x):

Config	TGFR Flow	TGFR DCN	Decoder	SAFR	LPIPS↓	tLPIPS↓	tOF↓	PSNR↑
(a) Baseline (StableVSR)			D		9.74	5.40	17.20	27.97
(b) + \(D_s\) w/o TGFR			\(D_s\)		9.65	5.45	17.24	28.30
(c) + Flow Alignment	✓		D		9.73	5.12	17.17	28.24
(d) + DCN Fine Alignment	✓	✓	D		9.71	4.95	17.05	28.35
(e) + \(D_s\) w/ TGFR	✓	✓	\(D_s\)		10.06	4.43	16.92	28.61
Ours (Full)	✓	✓	\(D_s\)	✓	9.54	4.20	16.79	28.75

Long-Sequence Gating Ablation (Table 4, REDS4 100 frames):

Variant	5×20 Frames LPIPS↓	5×20 Frames tOF↓	100 Frames Single LPIPS↓	100 Frames Single tOF↓
w/o Gate (Concat.)	12.59	27.28	42.65	1887.47
Gate w/o Sparsity	11.85	25.85	40.57	1718.06
Full TGFR	11.05	18.52	9.54	16.79

Key Findings

• TGFR is the primary factor in suppressing flickering: Without TGFR-based alignment and gating, the continuous decoder tends to amplify flickering. The two-stage alignment (Flow + DCN) effectively addresses feature drifts at different granularities.
• SAFR is key to balancing perception and fidelity: Directly mapping the continuous decoder to the aligned features (from (e) to Ours) causes a drop in perceptual quality due to inadequate high-frequency synthesis. SAFR restores these high-frequency details precisely through scale conditioning in the Fourier domain.
• Gated sparsity regularization is the cornerstone of long-sequence stability: Though seemingly a minor training detail, it determines whether the model can generalize to real-world long videos. Achieving uncorrupted single-pass inference over 100 frames is essential for practical deployment.
• Scale-agnostic denoising yields fundamental efficiency advantages: Peak memory stays constant around 8.7GB, and the computation scales up by only 1% from 2x to 8x. In comparison, VEnhancer experiences a 13.5x increase in FLOPs and a 1.8x spike in peak memory when moving from 2x to 8x.

Highlights & Insights

• Effective Decoupling: Decoupling is realized not only structurally (into two stages) but also computationally. Denoising diffusion sampling runs exclusively in the LR space, making generation cost entirely independent of output resolution. This design principle can be extended to any generative inference task requiring multi-resolution outputs (e.g., multi-resolution image generation, variable-resolution NeRF rendering).
• Gating Map Conditioned on the Absolute Difference \(|C-H|\) is a Masterstroke: Most feature fusion techniques rely solely on concatenation for gate prediction. AVSR-Diff explicitly introduces the absolute difference between the current and propagated features as an input to the gating network, enabling it to assess the local reliability of propagated information. This technique is highly reusable for any fusion scenario requiring the validation of external feature reliability.
• Elegant Scale Conditioning in the Frequency Domain: Instead of relying on brute-force spatial convolutions to adapt to diverse scales (which is computationally inefficient), SAFR uses a learnable spectral gating vector in the Fourier domain to directly control the energy of various frequency components. This approach is parameter-efficient, possesses clear physical interpretations (high frequencies for details, low frequencies for structures), and models the mapping to scale embeddings in an end-to-end manner.
• Pragmatic OOD Strategy (Training on \(s \in [1.1, 4.0]\) and Clipping to \(s=4\) for 8x Inference): Rather than designing complex scale-extrapolation mechanisms (e.g., relative positional encodings), simply clipping to the upper bound of the training scale while querying coordinates via LIIF is both simple and highly effective. This demonstrates that in the hybrid "diffusion + INR" paradigm, the division of labor—where generative priors are drawn from a robust 4x model and coordinate querying handles spatial mapping—inherently provides solid generalization capabilities.

Limitations & Future Work

• Slow Absolute Inference Speed: Generating a single frame from a \(180 \times 320\) input takes approximately 42 seconds (using 50-step DDPM), maintaining an order-of-magnitude latency gap compared to single-pass INR methods. The authors suggest that integrating faster sampling techniques such as DPM-Solver or consistency distillation presents an obvious and feasible direction for future work.
• Generative Prior Constrained by the 4x Base Model: Since the framework is built atop the SD×4 Upscaler, continuous decoding at larger scales (e.g., 8x) relies on extrapolating from 4x generative priors rather than exploiting stronger native generative capabilities. Utilizing generative foundation models trained across multiple native resolutions could unleash better large-scale performance.
• Dependence on Optical Flow: The temporal alignment pipeline (optical flow warping in TGFR, initial alignment in DCN) is fully reliant on RAFT. Consequently, performance may degrade in scenarios where optical flow estimation fails, such as fast motion, occlusions, or motion blur. The paper does not provide a systematic analysis of these failure modes.
• Gap Between Training and Inference Scales: The training scale is bounded at 4x, leaving 8x inference dependent on scale clipping and INR-based extrapolation. Performance under extreme upscaling factors (e.g., 16x, 32x) remains uninvestigated.

Related Work & Insights

• vs. StableVSR [30]: AVSR-Diff's temporal conditioning strategy (bidirectional sampling + warped RGB guidance) is directly inherited from StableVSR. However, StableVSR is restricted to a fixed 4x factor, as its decoder cannot perform arbitrary-scale rendering. AVSR-Diff improves upon it by introducing feature-level alignment via TGFR (as opposed to token/RGB-level warping), frequency-domain modulation using SAFR, and coordinate rendering through LIIF, successfully pushing the boundary of generative VSR from a fixed 4x factor to arbitrary, continuous scales.
• vs. VEnhancer [16]: Both methods target arbitrary-scale generative VSR but adopt opposite design philosophies. VEnhancer performs 3D U-Net denoising directly at the target resolution (heavy denoising at HR), while AVSR-Diff executes denoising in the LR space coupled with lightweight continuous decoding. Consequently, at 8x upscaling, AVSR-Diff is approximately 2.5x faster, reduces tOF by 3.3x, and consumes only 1/8 of the peak memory compared to VEnhancer. This split architecture—performing denoising in a low-dimensional space and rendering in a high-dimensional space—is highly recommended for wider adoption in other pixel-generation tasks.
• vs. INR Methods (e.g., LIIF, VideoINR): AVSR-Diff essentially replaces the direct regression from LR features to RGB in traditional INR methods with the generative priors of a diffusion model. While retaining the continuous coordinate querying mechanism of INRs (via LIIF rendering), the queried features are extracted from the diffusion-denoised latent space rather than deterministic encoder outputs. This validates the effectiveness of combining generative priors with continuous implicit representations for video tasks, provided that temporal feature alignment is strictly handled.

Rating

• Novelty: ⭐⭐⭐⭐☆ The decoupling framework, TGFR, and SAFR offer clear innovative motivations. However, as the underlying components (ControlNet, LIIF, DCN, optical flow warping) rely on existing technologies, the work is characterized as an "ingenious integration" rather than a brand-new paradigm.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ The main tables cover both fixed 4x and multi-scale upscaling (2x/3.25x/8x) against over 10 baselines. The ablation studies sequentially dissect the contributions of individual components. Additional tests, including 100-frame stability, scaling efficiency analyses, and OOD strategy comparisons, set a benchmark for evaluation in this domain.
• Writing Quality: ⭐⭐⭐⭐☆ The methodology is logically sound, and the comparisons and architectural overviews are highly compelling. The ablation narrative—charting a path where performance temporarily degrades before recovering (improving temporal metrics from (b) to (e) at the expense of perception, then achieving a dual victory with SAFR)—flows with excellent rhythm.
• Value: ⭐⭐⭐⭐⭐ This work represents the first practical realization of the hybrid "diffusion generation + arbitrary-scale continuous decoding" paradigm in the video domain. Its fixed-scale performance exceeds dedicated 4x methods, and its constant memory footprint makes it deployable on consumer-grade GPUs, demonstrating both solid engineering value and academic contribution.

Related Papers

• [CVPR 2026] STCDiT: Spatio-Temporally Consistent Diffusion Transformer for High-Quality Video Super-Resolution
• [ECCV 2024] Enhancing Perceptual Quality in Video Super-Resolution through Temporally-Consistent Detail Synthesis using Diffusion Models
• [ECCV 2026] DTI: Dynamic Trajectory Initialization for Generative Face Video Super-Resolution
• [CVPR 2026] Physics-Consistent Diffusion for Efficient Fluid Super-Resolution via Multiscale Residual Correction
• [ECCV 2024] You Only Need One Step: Fast Super-Resolution with Stable Diffusion via Scale Distillation