Allo{SR}2: Rectifying One-Step Super-Resolution to Stay Real via Allomorphic Generative Flows

Conference: ECCV 2026
Paper: Official page · PDF
Area: Real-World Image Super-Resolution / Flow Matching
Keywords: One-step super-resolution, SNR-guided initialization, trajectory consistency, allomorphic trajectory matching, generative prior

TL;DR

AlloSR2 inserts low-resolution latents at an intermediate point of a pretrained generative flow, uses FATC to straighten the local restoration trajectory and ATM to align it with a parameter-shared generative trajectory, and performs 4x super-resolution with one velocity prediction, improving several realism metrics without targeting the highest pixel fidelity.

Background & Motivation

Real-world super-resolution must handle combinations of blur, compression, and noise rather than a single bicubic downsampling operator. Generative models can supply textures that conventional regression networks tend to smooth away, but methods such as StableSR and SeeSR require repeated denoising. Compressing sampling into one step reduces that burden, while making it harder to preserve the pretrained model's knowledge of natural images during substantial task adaptation.

The paper targets a specific mismatch at initialization. A generative model expects Gaussian noise, whereas a restoration model often substitutes a degraded image latent at the same entry point. An LR image still contains substantial structure and is not pure noise. Supervising only the final HR output also leaves the intermediate velocity field unconstrained. The network can therefore learn a shortcut that fits training pairs but departs from the generative trajectory, producing over-smoothed textures or unnatural details. Here, "prior collapse" refers to the deterioration of generative capability during task-specific adaptation, not an established theorem about a particular probability distribution collapsing.

AlloSR2 treats generation and restoration as differently shaped paths toward natural HR images: a global path starts from noise, and a local path starts from a degraded observation. They share a velocity network, so the local path can support one-step integration while the global path supplies a continuing distributional constraint. Core idea: locate the observation at a suitable point on the generative time axis, then correct both local velocities and cross-trajectory distributions instead of merely fine-tuning a generative backbone into an endpoint regressor.

Method

Overall Architecture

The input is a degraded LR image and the output is a restored image at 4x resolution. During training, paired LR-HR latent statistics determine the entry time; FATC constrains the local SR flow, ATM aligns the local SR flow with the global generative flow in a shared velocity field, and endpoint reconstruction supervision preserves content corresponding to the input. Inference retains VAE encoding, one velocity prediction at the selected time, and VAE decoding. It does not sample both training trajectories.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    Input["LR-HR training pairs<br/>VAE latents"] --> Init["SNR-Guided<br/>Trajectory Initialization"]
    Init --> FATC["Flow-Anchored<br/>Trajectory Consistency<br/>FATC"]
    FATC --> ATM["Allomorphic<br/>Trajectory Matching<br/>ATM"]
    Global["Noise-to-HR generative flow<br/>Velocity network shared with SR"] --> ATM
    ATM --> Model["Trained one-step SR model"]
    LR["Inference: encode LR"] --> Model
    Model --> Output["One velocity prediction<br/>Euler update and decoding"]

The FATC-to-ATM connection organizes the training constraints; it does not denote two sequential network calls at inference. The paper does not specify a clearly separated two-stage optimization schedule, so it should not be rewritten as "train FATC first, then train ATM separately."

Key Designs

1. SNR-Guided Trajectory Initialization: enter the flow where the observation's information content belongs

Let \(z_H\) denote the HR latent and \(z_L\) the LR latent, with degradation residual \(\Delta z=z_L-z_H\). On the normalized time axis of a linear flow, \(t=0\) denotes clean data and \(t=1\) denotes noise. LR is neither an uninformative noise sample nor an already restored HR image, making an intermediate entry point more appropriate. The authors encode paired samples using the fixed pretrained VAE, characterize degraded SNR by the ratio of HR signal variance to residual variance, and search for the \(t^*\) whose theoretical trajectory SNR is closest. This is a dataset-statistics-based choice; the paper does not introduce a per-image degradation estimator that predicts adaptive entry times.

The choice skips the high-noise region beyond \(t^*\) and lets restoration start from available structure rather than requiring the network to interpret structured LR content as noise. However, SNR aligns a statistic, not necessarily the complete latent distribution, and initialization alone does not prove that the prior cannot be forgotten. The summation and expectation layout of cached Eq. (6) is damaged, and the text does not report the selected numerical \(t^*\). This note therefore retains the supported search logic without inventing an exact implementation. The experimental labels \(t=999\) and \(t=1\) refer to discrete scheduler indices; the latter must not be confused with the pure-noise endpoint of the normalized time axis.

2. Flow-Anchored Trajectory Consistency: supervise how the model moves, not only where it ends

A single update from the selected entry point approximates an entire integral with the initial velocity. Endpoint supervision alone can fit a complicated shortcut without producing a stable intermediate velocity field. FATC samples intermediate states between HR and LR, rescales the local path to the interval \([0,t^*]\), and supervises its velocity with the slope of that line. Following the prose accompanying Eqs. (7), (9), and (10), the central relations can be written as:

\[ \hat z_H=z_L-t^*v_\theta(z_L,t^*),\qquad z_t^{\mathrm{sr}}=\left(1-\frac{t}{t^*}\right)z_H+\frac{t}{t^*}z_L,\qquad \mathcal L_{\mathrm{FATC}}=\mathbb E\left[\left\|v_\theta(z_t^{\mathrm{sr}},t)-\frac{z_L-z_H}{t^*}\right\|_2^2\right]. \]

The minus sign comes from reverse-time integration: the velocity is defined toward LR, while restoration moves from \(t^*\) back to zero. Dividing by \(t^*\) is essential because the same latent displacement must be completed over a shortened time interval. FATC encourages consistent velocities along the path to reduce the error of a one-step Euler approximation, without adding inference steps. These equations normalize damaged formatting using the readable prose; they do not imply that every equation character survives in the cache. The sampling distribution and implementation details still require the original equations or code.

This supervision remains grounded in paired LR-HR data. It directly constrains the reconstruction path rather than the complete natural-image distribution. A straighter trajectory therefore does not imply that synthesized details must be realistic, which motivates the next design.

3. Allomorphic Trajectory Matching: constrain the SR output distribution with a shared generative flow

ATM addresses the loss of generative richness that can result from paired regression alone. Inspired by TwinFlow's self-adversarial formulation, the authors share parameters between the global noise-to-HR flow and the local LR-to-HR flow, matching scores at corresponding times within a unified velocity field. A score is the gradient of log probability density with respect to a sample: it indicates how to move toward higher density under that distribution, rather than being an additional image-quality rating. The global flow supplies dynamic positive guidance instead of serving only as a static pretrained backbone.

The distributional objective encourages restored latents to match the generative prior. The direction in Eq. (11) is:

\[ \mathcal L_{\mathrm{ATM}}=D_{\mathrm{KL}}\!\left(p_{\mathrm{sr}}\parallel p_{\mathrm{gen}}\right). \]

Because high-dimensional densities cannot be directly evaluated, training uses a score difference to supply gradients to the restored output. Eq. (12) connects this difference to two kinds of information: positional alignment between local and global intermediate states, and directional consistency between predicted velocities at the same time, with a time-dependent weight \(w(t)\). Positional alignment prevents similarly directed paths from remaining displaced in latent space; directional consistency prevents nearby states from subsequently moving apart. Together they provide cross-trajectory distributional guidance, rather than simply pulling the final SR image toward an unrelated randomly generated image.

The cached expansion of Eq. (12) is severely disordered, so its complete signs, scaling factors, and stop-gradient behavior cannot be verified reliably. The prose also does not fully specify \(w(t)\), the global branch's training objective, or the update order. This note therefore does not invent runnable ATM pseudocode or describe shared parameters as two independent teacher networks. The supported distinction is that FATC constrains the local trajectory for integration, whereas ATM preserves the generative character of the output distribution.

A Worked Example

Consider a \(128\times128\) degraded crop from the evaluation setup. The task is to produce a \(512\times512\) restored image, not to generate an unrelated scene from pure noise. The VAE maps the input to \(z_L\), the velocity network predicts once at the selected \(t^*\), the Euler update above yields \(\hat z_H\), and decoding produces the HR output. The cache does not spell out the exact resizing preprocessing before the encoder.

For training intuition, the paired HR crop supplies \(z_H\): FATC teaches stable movement through states between the endpoints, while ATM keeps this restoration trajectory aligned with the natural-image generative flow. The crop dimensions come from the experimental protocol; this walkthrough is not a new measured case study. One function evaluation means one velocity-network call, excluding VAE encoding and decoding. Tiled processing also does not mean that an entire large image needs only one network call.

Loss & Training

The endpoint reconstruction objective combines latent-space MSE with pixel-space LPIPS, comparing decoded predicted and HR latents for the latter. MSE preserves content corresponding to the input, LPIPS constrains perceptual similarity in feature space, and FATC and ATM add local-velocity and distributional constraints. The paper introduces an LPIPS coefficient \(\lambda\), but the cache does not supply its numerical value or a complete configuration of weights across the objectives.

The foundation model is FLUX.1-dev, with rank-64 LoRA added to both the VAE encoder and the DiT. SNR statistics use the fixed pretrained encoder, while later adaptation trains encoder LoRA; these are separately stated parts of the procedure. Whether SNR is re-estimated after adaptation is unspecified. Training uses LSDIR and the first 10K FFHQ face images, totaling 95K images, with paired samples synthesized through the Real-ESRGAN degradation pipeline.

Optimization uses AdamW, learning rate \(5\times10^{-5}\), global batch size 16, and 10K iterations. The initialization analysis separately reports convergence around 7K iterations with the selected time versus 9K with other settings. This does not replace the common training budget with 7K iterations and is not a complete hardware training-time comparison.

Key Experimental Results

Main Results

All experiments use 4x super-resolution. DIV2K-Val uses synthetic degradation; RealSR and DRealSR provide paired ground truth and use \(128\times128\) center-cropped LR inputs; RealLQ250 has no paired ground truth and retains \(256\times256\) inputs. The main configuration trains and evaluates at \(512\times512\), with tiled inference for larger images.

The following selection from Table 1 compares the one-step methods OSEDiff and TSD-SR without presenting them as the previous best on every metric. Higher PSNR, MUSIQ, and MANIQA are better; lower LPIPS and FID are better. FID compares feature distributions over image collections and is not a measure of an individual image's content fidelity.

Dataset	Metric	OSEDiff	TSD-SR	AlloSR2
DIV2K-Val	PSNR	23.72	23.02	22.98
DIV2K-Val	LPIPS	0.2941	0.2673	0.2854
DIV2K-Val	MUSIQ	67.97	71.69	72.00
RealSR	FID	123.49	114.45	112.88
RealSR	MANIQA	0.6326	0.6347	0.6725
DRealSR	FID	135.30	134.98	132.48
DRealSR	MUSIQ	64.65	66.62	66.95
RealLQ250	MUSIQ	69.55	72.09	72.51

RealSR FID improves over TSD-SR by 1.57, but DIV2K-Val LPIPS is worse than TSD-SR's. The full table also shows lower RealSR MUSIQ than TSD-SR and non-leading RealLQ250 NIQE and CLIPIQA. The evidence supports improvements in selected naturalness and distributional metrics, not a blanket lead on every perceptual measure.

Ablation Study

The following DRealSR loss ablation comes from Table 2 and preserves metrics whose changes disagree in direction.

Configuration	PSNR ↑	LPIPS ↓	DISTS ↓	FID ↓	MUSIQ ↑	MANIQA ↑
Full model	27.95	0.2989	0.2128	132.48	66.95	0.6264
Without FATC	27.88	0.3106	0.2105	121.75	63.49	0.6192
Without ATM	28.68	0.3008	0.2003	139.74	62.77	0.6049

Removing FATC lowers MUSIQ from 66.95 to 63.49 but improves FID from 132.48 to 121.75. Removing ATM lowers MUSIQ to 62.77 and worsens FID to 139.74 while improving PSNR and DISTS. The ablation must not be summarized as "every component improves every metric," nor should all perceptual measures be treated as the same objective.

Key Findings

• Initialization is not best on every metric: on RealSR, selected \(t^*\) gives PSNR 24.97, versus 24.16 at \(t=999\) and 24.74 at \(t=1\). However, \(t=1\) yields LPIPS 0.2741, slightly better than 0.2755 at the selected time.
• Higher native resolution helps: moving the RealLQ250 configuration from \(512\times512\) to \(1024\times1024\) raises MUSIQ from 72.51 to 72.93 and lowers NIQE from 3.59 to 3.52. Both training and evaluation configurations change, so this is not an isolated inference-resolution ablation.
• Efficiency evidence primarily concerns step counts: StableSR, SeeSR, and ResShift use 200, 50, and 15 steps, respectively, versus 1 here. A 15-to-200-fold reduction in function evaluations cannot be treated as the same end-to-end speedup.

Highlights & Insights

• Constraining the entry point, path, and distribution separately provides a clearer account of one-step restoration than endpoint optimization alone. It also suggests separately measuring initialization mismatch, trajectory curvature, and changes in output distributions.
• The important FATC detail is local time rescaling, not a generic consistency objective. Transferring the idea to denoising or deblurring would require incorporating the degradation entry interval into the velocity target as well.
• ATM makes the generative prior a dynamic training constraint, an idea worth exploring for other generative restoration tasks. This is a methodological opportunity, not a claim that the paper already evaluates those tasks.

Limitations & Future Work

• The authors discuss the perception-distortion trade-off but do not provide a separate systematic limitations section. Their ablations additionally show trade-offs between perceptual metrics themselves; FID or blind quality scores alone cannot establish correct content recovery.
• The cache does not report the selected numerical \(t^*\), full loss weights, ATM update details, or the time weighting, and some mathematical formatting is damaged. Reproduction still requires complete equations or the authors' implementation; the available information is not a directly reproducible training recipe.
• There is no complete efficiency comparison with matched backbones and resources, nor wall-clock or memory statistics. The cost of one FLUX.1-dev step cannot be inferred to be lower than every smaller multi-step model from call counts alone.
• Evaluation covers four datasets, but training degradation is primarily synthetic. Systematic extreme-degradation tests, text-accuracy evaluation, and face-identity preservation evaluation are absent. Degradation-stratified testing and content-fidelity metrics could distinguish plausible details from faithful ones.
• It remains unclear whether a dataset-level SNR entry time accommodates different degradations in individual images. Per-image or per-region time estimation is a future direction, not an existing capability of the method.

Related Work & Insights

• Versus OSEDiff / TSD-SR: These methods use variational score distillation and target score distillation, respectively, for one-step restoration. AlloSR2 emphasizes dual-trajectory alignment in a shared velocity field and additionally addresses initialization and the local path. Gains should be compared metric by metric.
• Versus CTMSR: CTMSR also addresses consistency and trajectory matching, so one-step trajectory supervision alone is not unique to this paper. The specific combination here is SNR-based entry, FATC velocity supervision, and ATM's allomorphic generative-flow constraint.
• Connections to OMGSR / TwinFlow: OMGSR provides context for intermediate-timestep guidance, while TwinFlow inspires self-adversarial shared-flow distillation. AlloSR2 places these ideas within an LR-content-constrained local restoration path; its contribution lies mainly in the specific restoration formulation and its combined evaluation.

Rating

• Novelty: 4/5. Three mechanisms address a concrete trajectory mismatch, though intermediate-time guidance and self-adversarial dual flows have precedents.
• Experimental Thoroughness: 3/5. Four test sets and multiple ablations are included, but resource-matched comparisons, content-fidelity evaluation, and complete reproduction details are missing.
• Writing Quality: 3/5. The motivation is coherent, but claims of optimality and some perceptual conclusions require reading the counterexamples; damaged cached equations separately reduce readability.
• Value: 4/5. The method provides a clear division of responsibilities for one-step generative restoration, especially when naturalness matters and fidelity trade-offs are acceptable.