OMGSR: Mid-timestep Guidance in Real-ISR

• The paper's main contribution is replacing initial-step latent injection with mid-timestep guidance that better aligns LQ image latents with pretrained generative priors.
• It introduces a Latent Distribution Refinement loss and overlap-chunked LPIPS/GAN losses to fine-tune latent matching and suppress artifacts in high-resolution outputs.
• OMGSR outperforms other one-step Real-ISR methods by achieving faster convergence and detailed restoration on benchmarks using DDPM and flow matching variants.

Searching arXiv for the OMGSR paper and closely related one-step Real-ISR work. One Mid-timestep Guidance Real-ISR (OMGSR) is a one-step framework for real-world image super-resolution (Real-ISR) that injects the low-quality (LQ) image latent distribution at a pre-computed mid-timestep rather than at the initial timestep of a diffusion or flow-matching process. The method is motivated by the observation that the noisy latent distribution at DDPM/FM mid-timesteps aligns more closely with the LQ image latent distribution, thereby reducing the latent distribution gap that arises when one-step Real-ISR models directly substitute an LQ latent for the Gaussian noisy latent expected by the pretrained generative model. OMGSR is presented as a universal framework applicable to DDPM/FM-based generative models, with two instantiated variants—OMGSR-S based on SD-Turbo and OMGSR-F based on FLUX.1-dev—and combines mid-timestep injection, a Latent Distribution Refinement loss, and Overlap-Chunked LPIPS/GAN losses for artifact suppression and high-resolution training (Wu et al., 11 Aug 2025).

1. Conceptual basis and problem setting

OMGSR is situated in the line of work that seeks to make Real-ISR efficient by collapsing diffusion-style generation into a single denoising step. In the formulation described for the method, recent one-step Real-ISR models typically inject an LQ image latent distribution at the initial timestep. The central claim of OMGSR is that this practice produces a fundamental gap between the LQ image latent distribution and the Gaussian noisy latent distribution used during pre-training, which in turn limits the effective utilization of generative priors (Wu et al., 11 Aug 2025).

The paper’s key empirical insight is that the LQ image latent distribution is much more similar to the model’s latent at a mid-timestep than at the initial, full-noise state. This observation reorients one-step Real-ISR from initial-timestep substitution toward what the paper calls “one mid-timestep guidance.” In this design, the LQ latent is not treated as a proxy for pure noise; instead, it is matched to a location on the pretrained generative trajectory where its statistics are closer to those encountered by the backbone during pre-training.

This framing distinguishes OMGSR from deterministic one-step methods such as OSEDiff, which directly take the LQ image as the starting point and eliminate random noise, and from later time-aware or controllable one-step methods that learn timestep-dependent or dual-stream control mechanisms (Wu et al., 2024, Zhang et al., 22 Aug 2025, Fang et al., 21 Nov 2025). A plausible implication is that OMGSR occupies an intermediate position between strict restoration and generative super-resolution: it retains a one-step computational profile while preserving closer adherence to the pretrained model’s latent-space operating regime.

2. Mid-timestep matching and latent distribution refinement

The operational core of OMGSR is the selection of a pre-computed optimal mid-timestep $t_m$. The paper defines the optimal mid-timestep by minimizing the average $L_2$ distance between the LQ latent $z_L$ and the model’s pretrained noisy latent $z_{t_m}$ over a dataset of LQ-HQ pairs:

$$t_m^* = \mathrm{argmin}_{t_m} \frac{1}{D} \sum_{i=1}^D \| z_L^{(i)} - z_{t_m}^{(i)} \|_2^2$$

where $D$ is the dataset size (Wu et al., 11 Aug 2025).

In practice, the reported optimal values are approximately $t_m \approx 195$ for SD-Turbo and $t_m \approx 295$ for FLUX.1-dev. These values are not introduced as generic constants for all backbones; they are reported for the two instantiated OMGSR variants. The method uses these mid-timesteps during both training and inference, so the LQ image latent is input at $t_m$ rather than at the initial step.

The paper further states that even at $t_m$ there remains a slight mismatch, and therefore introduces the Latent Distribution Refinement loss, denoted $L_2$0. Its role is to fine-tune the VAE encoder for LQ images, using LoRA, so that the latent $L_2$1 better matches the corresponding pretrained noisy latent generated from the high-quality latent $L_2$2 and sampled noise $L_2$3. The two backbone-specific forms are:

For DDPM

$L_2$4

For FM

$L_2$5

This refinement mechanism is central to OMGSR’s claim of “bridging the latent distribution gap.” A common misconception in one-step Real-ISR is that the main difficulty is only insufficient denoising capacity in a single forward pass. OMGSR instead attributes a substantial part of the failure mode to latent-space misalignment between the LQ encoder output and the pretrained model’s trajectory.

3. Inference mechanism and model formulations

During training and inference, OMGSR performs a denoising prediction step from the selected mid-timestep and decodes the resulting latent with the VAE decoder. The framework is designed to be universal across DDPM-based and FM-based generative models, and the paper gives separate prediction rules for each regime (Wu et al., 11 Aug 2025).

For DDPM-based instantiations, the inversion step is written as:

$L_2$6

where $L_2$7 is the prompt or condition and $L_2$8 is the UNet/DiT prediction. For FM-based instantiations, the prediction is:

$L_2$9

and the output image is then recovered as

$z_L$0

with $z_L$1 denoting the VAE decoder.

The method therefore preserves the pretrained denoising backbone and conditioning pathway, but relocates the point at which the LQ latent enters the generation process. This suggests that OMGSR does not merely modify the loss; it changes the geometric relation between the restoration input and the pretrained generative trajectory. The paper’s terminology of “guidance” refers to this injection at the selected mid-timestep, rather than to multi-step guidance in the conventional iterative diffusion sense.

4. Overlap-chunked losses and high-resolution training

A second technical component of OMGSR is the Overlap-Chunked LPIPS/GAN loss, abbreviated as OC-LPIPS/GAN. The paper motivates this design by noting that standard perceptual and GAN losses require low-resolution inputs, such as 224–518 pixels for VGG/DINOv2, and that naive resizing of high-resolution images such as $z_L$2 or $z_L$3 causes severe checkerboard artifacts (Wu et al., 11 Aug 2025).

To address this, OMGSR splits images into overlapping patches and computes perceptual or adversarial losses patchwise. The LPIPS form is

$z_L$4

where $z_L$5 extracts the $z_L$6th chunk or patch and $z_L$7 is the number of patches. The GAN counterpart is formulated similarly using a patch-based DINOv2 discriminator:

$z_L$8

The full objective is given as

$z_L$9

with the $z_{t_m}$0 coefficients empirically set in the paper. Within the internal logic of the framework, OC-LPIPS/GAN serves a different role from mid-timestep guidance: the latter addresses latent distribution alignment, whereas the former addresses artifact suppression and the practical problem of optimizing very high-resolution outputs without destructive resizing. The paper explicitly reports that without OC-LPIPS/GAN, checkerboard artifacts appear in all high-resolution settings.

5. Variants, resolutions, and empirical findings

OMGSR is instantiated in two variants, one for DDPM and one for flow matching. The reported variants are summarized below (Wu et al., 11 Aug 2025).

Variant	Backbone	Best $z_{t_m}$1
OMGSR-S	SD-Turbo (DDPM)	195
OMGSR-F	FLUX.1-dev (Flow Matching, larger, 1k)	295

The main reported evaluation resolution is $z_{t_m}$2, described as “4×” super-resolution from 128 or 256 inputs, with additional experiments at $z_{t_m}$3 and $z_{t_m}$4 via two-stage upscaling with OMGSR-F. The metrics listed in the paper include PSNR and SSIM for fidelity, and LPIPS, DISTS, FID, NIQE, MUSIQ, M-IQA, and C-IQA for perceptual or no-reference evaluation (Wu et al., 11 Aug 2025).

The paper states that on the major datasets RealSR, DrealSR, and DIV2K, OMGSR-F dominates in all reference metrics, outperforming state-of-the-art one-step methods such as OSEDiff, SinSR, PiSA-SR, TSD-SR, and FluxSR, as well as multi-step methods such as StableSR, DiffBIR, and SeeSR. OMGSR-S is reported to achieve the best NIQE among the tested methods. For qualitative performance, the paper highlights that 1k-OMGSR-F restores fine text, specifically “eduroam,” and small facial details where all other models fail, and that 2k-generation by sequential $z_{t_m}$5 and $z_{t_m}$6 upscaling yields detailed, artifact-free high-resolution images.

The paper further reports that OMGSR-S is the fastest among one-step methods, whereas OMGSR-F is slower because of its model size but still consists of three steps—VAE encode, denoise, and VAE decode. In ablation, using $z_{t_m}$7 or $z_{t_m}$8 gives substantially better details and faster convergence than $z_{t_m}$9, which is identified with legacy one-step distillation.

6. Position within the one-step Real-ISR literature

OMGSR belongs to a rapidly developing family of one-step Real-ISR methods, but its defining intervention is more specific than the generic term “one-step diffusion.” OSEDiff formulates one-step Real-ISR by directly setting the latent of the input LQ image as the starting point for diffusion and applying variational score distillation in latent space, thereby emphasizing determinism and elimination of random noise (Wu et al., 2024). OMGSR instead retains a noised generative trajectory and argues that the correct entry point is a mid-timestep rather than the initial timestep.

Subsequent methods elaborate related concerns from different angles. TADSR argues that a fixed timestep is suboptimal because Stable Diffusion exhibits different generative priors at different noise injection timesteps, and proposes a Time-Aware VAE Encoder plus a Time-Aware VSD loss so that a user can change the timestep condition to obtain controllable trade-offs between fidelity and realism (Zhang et al., 22 Aug 2025). IDaS-SR similarly critiques rigid timestep initialization, but replaces the pre-computed mid-timestep of OMGSR with a Manifold Inversion Noise Estimator that predicts a severity-aware timestep and inversion noise for each instance, together with CHARIOT for continuous steering along the perception–distortion boundary (Weng et al., 27 Apr 2026). ODTSR moves the problem into a diffusion-transformer setting and introduces a Noise-hybrid Visual Stream design and Fidelity-aware Adversarial Training so that a Fidelity Weight controls the trade-off between restoration and prompt-guided generation (Fang et al., 21 Nov 2025).

Taken together, these comparisons indicate that OMGSR’s historical significance lies in making mid-timestep injection a first-class design principle for one-step Real-ISR. A plausible implication is that later methods can be read as attempts to generalize, individualize, or make controllable the alignment problem that OMGSR framed at the level of a pre-computed mid-timestep. Within that trajectory, OMGSR remains defined by three linked claims: the latent distribution gap is central, a mid-timestep offers a better alignment point than the initial timestep, and high-resolution one-step training requires overlap-chunked perceptual and adversarial losses to avoid checkerboard artifacts (Wu et al., 11 Aug 2025).