AdaBridge-SR: Adaptive Bridge Matching for Real-World Image Super-Resolution

Conference: ECCV 2026
Paper: Official page / PDF
Code: https://github.com/W-JG/AdaBridge-SR
Area: Image Super-Resolution / Image Generation
Keywords: Bridge matching, spatial stochasticity, time reparameterization, latent restoration, perception-distortion trade-off

TL;DR

AdaBridge-SR predicts where to inject noise and when to enter high-variance bridge states from the degraded input, using shared weights for deterministic one-step super-resolution and stochastic few-step refinement: on RealSR, it reaches 25.2152 dB PSNR in one step and 0.6847 MANIQA in two steps, while additional steps still incur a fidelity cost.

Background & Motivation

Real-world image super-resolution is not simply about sharpening every location. An image may contain readable text contours, flat backgrounds, and a brick wall whose details have largely disappeared. The first two favor preservation of the observed structure; the last may need a generative prior to supply plausible texture. Multi-step diffusion models can synthesize details, but spatially uniform noise can disturb structures that were already reliable. One-step methods compress restoration into a direct mapping, reducing latency but also reducing stochastic exploration of ambiguous textures.

Bridge matching connects degraded and clean images as the two endpoints of generation, making restoration more direct than starting from pure noise. A standard Brownian bridge nevertheless uses a common noise amplitude and a fixed temporal schedule: every location receives the same perturbation scale, and variance peaks at the same time. Beyond spatial heterogeneity, overall degradation severity differs between images. Spatial weighting alone cannot determine when a particular input should receive more generative capacity during restoration.

The paper therefore separates stochasticity into spatial and temporal controls conditioned on the input, rather than appending a sharpening stage. Core idea: let the degraded latent determine both local noise amplitude and bridge time reparameterization, then use clean-endpoint prediction to unify one-step restoration and stochastic few-step refinement, allocating generative capacity according to location and degradation severity.

Method

Overall Architecture

A VAE first encodes the low-quality image into a degraded latent. The Adaptive Bridge Controller, ABC, predicts a spatial stochasticity map and a temporal warping exponent. The spatio-temporal adaptive bridge, ST-ABM, constructs an intermediate latent between degraded and clean endpoints with locally modulated noise. A denoising backbone predicts the clean endpoint directly, and the VAE decodes that prediction into an image.

The important training-inference distinction is the source of the clean endpoint. Training uses paired ground truth. Inference begins with a deterministic prediction, then treats the preceding prediction as a dynamic clean endpoint. One-step mode returns the initial prediction without evaluating ABC; few-step mode computes the controls and reuses them in subsequent steps. Supporting both modes with one model therefore does not mean that one-step inference also executes stochastic bridge refinement.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    Input["Degraded image<br/>VAE encoding"] --> Controller["Adaptive Bridge Controller"]
    Controller --> Bridge["Spatio-Temporal Adaptive Bridge"]
    Target["Training: ground-truth latent"] --> Bridge
    Bridge --> Endpoint["Endpoint Prediction and<br/>Dynamic Refinement"]
    Input -->|Direct prediction in one-step mode| Endpoint
    Endpoint -->|Few-step inference: previous endpoint| Bridge
    Endpoint --> Output["VAE decoding<br/>High-resolution image"]

Key Designs

1. Adaptive Bridge Controller: predict spatial and temporal controls from the same degraded input

ABC is a lightweight ResNet-based network that takes the degraded latent and uses two heads to predict the spatial map \(M\) and temporal exponent \(p\). Each spatial location has one weight, broadcast across latent channels: smaller weights suppress random perturbations, while larger weights permit more texture exploration. The map head applies a sigmoid, mean normalization, and clamping to stabilize its scale and avoid degenerate all-zero or all-one outputs. A bounded sigmoid-affine transformation restricts the temporal exponent to a positive interval. The available body text does not specify the interval endpoints or all implementation constants for map normalization; these cannot be filled in from the diagram alone.

Temporal control does not dynamically increase the number of network evaluations according to content. Instead, it changes where a fixed set of steps lands on the bridge. Given a base time fraction \(s\in[0,1]\), intrinsic time becomes \(\tau=s^p\). For \(p<1\), the same base time corresponds to a larger intrinsic time, moving the high-variance portion earlier. For \(p>1\), it moves later. The authors use this behavior to activate exploration earlier for severe degradation and postpone refinement for mild degradation. This controls trajectory positions and sampling allocation, rather than directly predicting the optimal step count for each image.

2. Spatio-Temporal Adaptive Bridge: explicitly separate local perturbation from temporal allocation

A conventional bridge linearly interpolates between endpoints and adds Gaussian perturbations that are stronger in the middle and vanish at both ends. The proposed bridge preserves this endpoint consistency but replaces uniform noise with spatially weighted noise and fixed time with the controller's intrinsic time. Let \(z_0,z_1\) denote degraded and clean latents, and let \(\beta\) be the global noise amplitude. Its central training relation is:

\[ \tau=s^p,\qquad z_\tau=(1-\tau)z_0+\tau z_1+\beta M\odot\sqrt{\tau(1-\tau)}\,\epsilon, \qquad \epsilon\sim\mathcal N(0,I). \]

This corresponds to Eq. (13). The cached extraction damages some plus signs, brackets, and equation endings. The notation above is normalized only from the adjacent text's explicit description of endpoint interpolation plus spatially modulated Gaussian noise; it does not supply missing hyperparameters. Refer to the original PDF for exact typesetting. Noise vanishes at both endpoints, preserving the bridge's start and end constraints. At a fixed time, conditional noise variance at a location is proportional to \(M^2\), not \(M\), because the map controls amplitude.

Where the spatial map approaches zero, the random component disappears; larger map values permit local stochastic exploration. Time reparameterization changes both the interpolation coefficient and the placement of the noise envelope, so it should not be understood as merely changing noise strength. The paper explains this through a bridge SDE in intrinsic time, followed by a time change that rescales both drift and diffusion. This continuous-time interpretation motivates the model, but its practical few-step algorithm uses the endpoint-reconstruction recurrence below rather than directly applying a generic SDE solver.

3. Endpoint Prediction and Dynamic Refinement: predict a complete result, then rebuild the next bridge from that prediction

The backbone outputs a clean-endpoint estimate rather than a velocity field that must subsequently be integrated. During training, the discrete set of sampled base times includes \(s=0\). At that time the bridge latent equals the degraded latent, allowing the same model to learn direct restoration from the input. For an inference budget of \(N=1\), the model performs only this endpoint prediction and decoding, requiring neither sampled noise nor controller outputs.

For \(N>1\), ABC runs once. At step \(k=2,\ldots,N\), the method sets \(s_k=(k-1)/N\) and \(\tau_k=s_k^p\). It replaces the ground-truth endpoint in the training bridge with the preceding prediction \(\hat z_1^{(k-1)}\), retains the original degraded latent as the other endpoint, and adds Gaussian noise for the current step. The backbone processes this reconstructed intermediate latent and predicts another complete endpoint. Each iteration thus returns to the original input as an anchor, instead of merely adding noise to the last generated image. However, the predicted endpoint can carry errors, so anchoring does not guarantee hallucination-free restoration.

A Worked Example

Consider an input containing recognizable sign edges and blurred wall texture. This is an illustration of the mechanism, not an additional experimental sample. One-step mode encodes the input, predicts the clean latent once, decodes it, and stops. Two-step mode keeps the first prediction and additionally evaluates ABC. Its second-step base time is \(s_2=1/2\), giving an actual mixing fraction of \((1/2)^p\). If the sign receives smaller map weights, it undergoes weaker perturbations, while larger weights on the wall allow more stochastic texture exploration.

If the whole image is severely degraded, a smaller \(p\) places the second step closer to the current predicted endpoint. With milder degradation, a larger \(p\) places the same step closer to the original input. The backbone then predicts a complete clean endpoint again, rather than returning the noisy intermediate as the final image. Specific spatial weights and exponents must come from the network; this example does not invent their values.

Loss & Training

The controller does not discover reliability entirely without supervision. Spatial supervision first computes the channel-averaged absolute difference between degraded and ground-truth latents at each position. It divides that residual map by its spatial mean plus \(\delta\), clips the result to \([0,1]\), and supervises \(M\) with SmoothL1, using \(\delta=10^{-6}\). This emphasizes relative difficulty within an image rather than treating absolute errors from different inputs as labels on a shared scale. Residuals above the mean are clipped, so the target is not a finely ranked probability estimate for all high-uncertainty regions.

Temporal supervision separately computes the global mean absolute latent residual and maintains an EMA of batch-mean residuals. The current sample's residual is divided by the EMA plus \(\delta\) and clipped to \([0,2]\) to obtain relative degradation severity. This is mapped linearly and inversely into the allowed exponent interval: greater severity produces a smaller target exponent. SmoothL1 supervises the predicted exponent. Paired clean latents are needed only during training to construct these targets; at inference the controller sees only the degraded input.

The total objective combines restoration loss with weighted spatial-map and temporal losses. Restoration supervision includes pixel reconstruction, perceptual, and adversarial terms, aligned with OMGSR. The cached body text does not provide the two controller-loss weights, EMA decay, or complete restoration-loss coefficients, so these are not presented as a reproducible configuration.

The model uses SD2.1-base and trains LoRA adapters in the VAE encoder and U-Net, with ranks 16 and 32 respectively, jointly with ABC. Training uses LSDIR and Real-ESRGAN's complex synthetic degradation, AdamW with learning rate \(5\times10^{-5}\), batch size 1, and gradient accumulation of 4, for 8000 steps on four RTX 4090 GPUs. The body text does not elaborate on the distributed meaning of this batch size, so no effective global batch size is inferred.

Key Experimental Results

Main Results

Evaluation uses fourfold super-resolution, following the StableSR protocol for RealSR, DRealSR, and synthetic DIV2K-Val. PSNR and SSIM are computed on the Y channel of YCbCr; LPIPS measures reference-based perceptual distance. MUSIQ, MANIQA, and CLIPIQA+ are no-reference quality metrics. The following clear rows are selected from Table 1. Higher PSNR and MANIQA are better; lower LPIPS is better.

Dataset	Method	Steps	PSNR / dB	LPIPS	MANIQA
RealSR	SeeSR	50	25.1478	0.3007	0.6450
RealSR	OSEDiff	1	25.1511	0.2920	0.6335
RealSR	Ours-S1	1	25.2152	0.2818	0.6546
RealSR	Ours-S2	2	24.8815	0.2881	0.6847
DRealSR	OSEDiff	1	27.9220	0.2967	0.5898
DRealSR	Ours-S1	1	28.2226	0.2969	0.6073
DRealSR	Ours-S2	2	27.8302	0.3103	0.6436
DIV2K-Val	OSEDiff	1	23.7235	0.2942	0.6131
DIV2K-Val	Ours-S1	1	23.9717	0.2926	0.6210
DIV2K-Val	Ours-S2	2	23.5063	0.2954	0.6556

The one-step model has the highest PSNR among the Table 1 comparisons on these three datasets, but does not lead every metric: its DRealSR LPIPS is slightly worse than OSEDiff's. On RealSR, moving from one step to two raises MANIQA from 0.6546 to 0.6847, while PSNR falls from 25.2152 to 24.8815 and LPIPS rises from 0.2818 to 0.2881. The method offers adjustable operating points within one model rather than eliminating the perception-distortion trade-off.

RealLQ250 has no clean references, so Table 2 evaluates only no-reference quality. Ours-S2 achieves 0.6218 MANIQA, above HYPIR's 0.6075, but its CLIPIQA+ of 0.7071 is below HYPIR's 0.7195. These results cannot establish that synthesized textures match the actual scene details.

Ablation Study

The following values come from Table 3, with two-step inference for every configuration. Spatial control corresponds to \(M\), and temporal control to \(p\). Disabling both means uniform noise and a fixed temporal schedule.

Spatial Control	Temporal Control	RealSR PSNR	RealSR LPIPS	RealSR MANIQA	DRealSR PSNR	DRealSR LPIPS	DRealSR MANIQA
Off	Off	24.6818	0.2893	0.6853	27.4918	0.3122	0.6424
Off	On	24.5748	0.3007	0.6886	27.4454	0.3375	0.6503
On	Off	24.6160	0.2906	0.6873	27.4471	0.3244	0.6463
On	On	24.8815	0.2881	0.6847	27.8302	0.3103	0.6436

Joint control gives the highest PSNR and lowest LPIPS among these four configurations on both datasets, but not the highest MANIQA. Temporal control alone worsens DRealSR LPIPS from 0.3122 to 0.3375; combining it with spatial control brings LPIPS down to 0.3103. The evidence therefore supports a better fidelity-perception compromise from joint control, not independent all-metric improvements from each module. Tables 1 and 3 differ slightly in the last reported digits of the full model's RealSR SSIM and MUSIQ; this note does not mix those columns.

Key Findings

• The step analysis covers \(N=1\) through \(4\): RealSR LPIPS changes from 0.2818 to 0.3138, while MUSIQ rises from approximately 67.62 to approximately 70.28 in the plot. More evaluations improve no-reference perceptual scores while worsening reference-based distance.
• Table 4 studies global noise amplitude under four-step inference. With \(\beta=0\), RealSR / DRealSR LPIPS is 0.3284 / 0.3744; with the adopted \(\beta=0.005\), it is 0.3138 / 0.3548. Moderate noise outperforms no noise here, but this four-step result must not be presented as a two-step ablation.
• Table 5 reports RTX 4090 latency at \(512\times512\): 96.38 ms per image for one step and 162.49 ms for two steps. ABC takes only 1.77 ms, or 1.09% of the two-step total. The controller is inexpensive; the second backbone evaluation is not free.

Highlights & Insights

• Separating where to generate from when to generate gives spatial heterogeneity and image-level degradation severity their own supervised mechanisms. Compared with simply masking noise, time reparameterization also changes the interpolation position of intermediate samples.
• Clean-endpoint prediction connects direct restoration and iterative refinement through the same parameterization. Training includes zero-time inputs, while few-step inference reuses predicted endpoints, avoiding separate weights for different budgets.
• The ablation shows that adding either control in isolation can move metrics in an unfavorable direction. A transferable experimental lesson is to report module interactions, reference errors, and no-reference quality together rather than selecting only the most favorable perceptual score.

Limitations & Future Work

• The paper does not provide a systematic failure-case or limitations section. The assessments below follow from its method and experimental boundaries; they are not presented as conclusions already verified by the authors.
• Latent residuals provide a usable difficulty proxy, not calibrated uncertainty. They may mix texture complexity, encoding error, and degradation strength. Mean normalization followed by clipping also suppresses differences among high-residual regions; correlation with actual restoration errors deserves direct testing.
• Training relies on LSDIR with synthetic degradation, while RealLQ250 has only no-reference metrics. Out-of-distribution degradations, text accuracy, and output stability across repeated stochastic samples need more direct evaluation.
• The continuous-time bridge supplies a theoretical interpretation, but inference uses a discrete recurrence with predicted endpoints. The body text does not guarantee exact sampling from the target distribution after this substitution; endpoint constraints are not guarantees of recovering true details.
• Controller hyperparameters, EMA settings, and loss coefficients are incompletely disclosed in the available cached body text, and the component table lacks repeated-run variance. The code URL comes from the paper; its accessibility and agreement with the description were not checked online for this note.

Related Work & Insights

• Compared with SeeSR, DiffBIR, and SUPIR: these multi-step diffusion restoration methods use generative priors and conditioning. AdaBridge-SR directly connects degraded and clean distributions while restricting local stochasticity with a spatial map. Fewer steps are an explicit advantage, but step counts across different systems do not translate directly into proportional speedups.
• Compared with OSEDiff and HYPIR: one-step approaches emphasize low-latency direct mappings. AdaBridge-SR retains a one-step entry point and adds optional dynamic bridge refinement for a higher-perceptual-score operating point. The reported values do not support simultaneous dominance on every metric.
• Compared with ResShift and UPSR: ResShift also transports degraded distributions toward clean ones. According to the paper's discussion, UPSR already uses spatial uncertainty to modulate stochasticity, but retains a fixed temporal schedule. AdaBridge-SR differs by jointly learning input-dependent temporal and spatial controls, not by being the first restoration method to use local stochasticity.
• A testable extension: for restoration where text or structural accuracy matters, report errors specifically in those regions to test whether the control map genuinely identifies reliable structures. This is a future research suggestion, not an experiment already completed in the paper.

Rating

• Novelty: 4/5. The joint design of spatial stochasticity and input-conditioned time reparameterization is clear, but bridge matching and local noise control have precedents.
• Experimental Thoroughness: 4/5. Four test sets, component interactions, steps, noise amplitude, and latency are covered; sampling variance and more direct semantic-fidelity evaluation remain missing.
• Writing Quality: 4/5. The theory-to-implementation narrative is clear, but several training details cannot be recovered from the current cache, and a few last-digit table values disagree.
• Value: 4/5. One model offers both low latency and an optional higher-perceptual-quality mode, providing practical flexibility rather than a cost-free improvement on every metric.