Realism Controlled One-step Diffusion (RCOD)

• The paper introduces RCOD, which reintroduces explicit control over the fidelity–realism trade-off by using the timestep as a control variable in one-step diffusion models.
• RCOD employs latent domain grouping and degradation-aware sampling to assign dynamic timesteps based on degradation metrics, ensuring tailored restoration outcomes.
• Empirical results demonstrate that RCOD achieves competitive PSNR, SSIM, and perceptual metrics while maintaining the efficiency of one-step inference.

Searching arXiv for the cited RCOD and closely related one-step diffusion restoration papers. Realism Controlled One-step Diffusion (RCOD) is a framework for real-world image super-resolution (Real-ISR) that reintroduces explicit control over the fidelity–realism trade-off into one-step diffusion models. In the formulation introduced in "Realism Control One-step Diffusion for Real-World Image Super-Resolution" (Wu et al., 12 Sep 2025), RCOD addresses a central limitation of one-step diffusion (OSD): although OSD methods are efficient, they are usually trained or distilled by a single timestep and therefore tend to converge to a single restoration regime. RCOD makes the timestep itself an explicit control variable, so that lower timesteps emphasize fidelity and higher timesteps emphasize realism, while preserving one-step inference efficiency (Wu et al., 12 Sep 2025).

1. Conceptual position within one-step diffusion restoration

RCOD emerges from the observation that diffusion-based real-world super-resolution has two competing objectives. The first is fidelity, namely preserving structure, content, and low-level details that remain in the low-resolution input. The second is realism, namely synthesizing perceptually convincing high-frequency textures and natural image statistics. In multi-step diffusion methods, this balance can be adjusted by changing the sampling trajectory or the number of denoising steps; in one-step diffusion methods, that flexibility is typically absent because the model is optimized for a single denoising configuration (Wu et al., 12 Sep 2025).

The explicit contribution of RCOD is to recover this missing control axis without abandoning the computational advantages of one-step inference. The framework is therefore best understood as a controllable one-step diffusion paradigm for Real-ISR, in which different noise levels correspond to different operating regimes during restoration (Wu et al., 12 Sep 2025).

This same fidelity-versus-realism tension also appears in adjacent one-step restoration research. CODSR, proposed in "Bridging Fidelity-Reality with Controllable One-Step Diffusion for Image Super-Resolution" (Chen et al., 16 Dec 2025), is a controllable one-step diffusion network for image super-resolution designed to balance “fidelity” and “reality” through LQ-guided modulation, region-adaptive prior activation, and text-matching guidance. HonestFace, proposed in "HonestFace: Towards Honest Face Restoration with One-Step Diffusion Model" (Wang et al., 24 May 2025), targets face restoration with an emphasis on identity consistency and texture realism. These methods do not define RCOD in the same way, but they illustrate a broader research direction in which one-step diffusion is augmented with explicit mechanisms that regulate hallucination, semantic conditioning, and realism.

2. Core formulation: timestep as a realism control variable

The foundational idea in RCOD is that the timestep controls the amount of noise in latent diffusion and, consequently, the degree of generative freedom available during denoising. The paper gives the forward process as

$$q(\mathbf{z}_t \mid \mathbf{z}_{t-1}) = \mathcal{N}(\mathbf{z}_t; \sqrt{1-\beta_t}\mathbf{z}_{t-1}, \beta_t \mathcal{I})$$

and

$$\mathbf{z}_t = \alpha_t \mathbf{z}_0 + \sigma_t \boldsymbol{\epsilon}, \quad \alpha_t = \prod_{s=1}^t \sqrt{1-\beta_s}, \quad \sigma_t = \sqrt{1-\alpha_t^2}.$$

Within this formulation, a larger timestep implies more injected noise, a larger deviation from the original latent, and correspondingly more scope for the denoiser to generate realistic but potentially less faithful detail (Wu et al., 12 Sep 2025).

RCOD operationalizes this by replacing the standard fixed-timestep one-step training setup with a grouped-timestep regime. The one-step restoration mapping is written as

$$\hat{\mathbf{z}}_H = F_\theta(\mathbf{z}_L; t, c_y),$$

and the selected timestep is no longer constant across training samples. Instead, it is assigned according to a latent-domain degradation metric. In the implementation reported in (Wu et al., 12 Sep 2025), the main experiments use three realism levels:

• Fidelity: $t=250$
• Neutral: $t=500$
• Realism: $t=750$

This discrete schedule creates a monotonic control axis in which increasing timestep increases realism (Wu et al., 12 Sep 2025). The paper’s reported numerical and visual results show the corresponding trend: the Fidelity setting improves PSNR and SSIM, the Neutral setting is intermediate, and the Realism setting improves no-reference perceptual metrics such as NIQE, MUSIQ, and CLIPIQA (Wu et al., 12 Sep 2025).

A plausible implication is that RCOD does not treat realism control as an auxiliary post hoc adjustment, but as a property learned directly in the denoiser’s latent operating modes.

3. Latent Domain Grouping and Degradation-Aware Sampling

The central mechanism that enables RCOD is Latent Domain Grouping (LDG). Rather than exposing all training samples to the same denoising condition, LDG partitions samples according to degradation level in latent space and maps them to different timestep groups. The latent degradation metric is defined as cosine similarity between the low-resolution and high-resolution latents:

$$M_L = \frac{\mathbf{z}_L \cdot \mathbf{z}_H}{\|\mathbf{z}_L\|\|\mathbf{z}_H\|}.$$

This metric is used only during training because $\mathbf{z}_H$ is available there. The grouping rule is given as

$$t = k \cdot \left(n - \left\lfloor \frac{n \cdot (M_L - M_{L\text{-min}})}{M_{L\text{-max}} - M_{L\text{-min}}} \right\rfloor \right), \quad k \in \mathbb{Z}^+.$$

In the reported implementation, the authors set $n \le 4$ and $$\mathbf{z}_t = \alpha_t \mathbf{z}_0 + \sigma_t \boldsymbol{\epsilon}, \quad \alpha_t = \prod_{s=1}^t \sqrt{1-\beta_s}, \quad \sigma_t = \sqrt{1-\alpha_t^2}.$$0 (Wu et al., 12 Sep 2025). The purpose of this grouping is to prevent the one-step model from collapsing into a single latent-domain behavior. Instead, the denoiser learns multiple restoration regimes aligned with degradation severity and desired perceptual outcome.

RCOD complements LDG with Degradation-Aware Sampling (DAS), which modifies the distillation regularization process so that the regularization timestep is sampled near the LDG-assigned timestep:

$$\mathbf{z}_t = \alpha_t \mathbf{z}_0 + \sigma_t \boldsymbol{\epsilon}, \quad \alpha_t = \prod_{s=1}^t \sqrt{1-\beta_s}, \quad \sigma_t = \sqrt{1-\alpha_t^2}.$$1

The stated purpose is to align distillation regularization with the grouping strategy, so that the model’s training signal remains locally consistent with the realism regime assigned to each sample (Wu et al., 12 Sep 2025). This is important because prior VSD-style one-step SR distillation methods sample regularization timesteps broadly, whereas RCOD requires regularization to respect the same degradation-conditioned structure that defines its control axis.

This suggests that RCOD’s controllability is not solely an inference-time parameterization. It is encoded into the training dynamics through the joint action of LDG and DAS.

4. Conditioning design and inference control

A second major component of RCOD is the Visual Prompt Injection Module (VPIM). In contrast to Stable Diffusion-based super-resolution methods that use text prompts or a vision-LLM for semantic guidance, RCOD replaces conventional text prompts with degradation-aware visual tokens extracted directly from the low-resolution image (Wu et al., 12 Sep 2025). The module uses a CLIP vision encoder followed by an MLP for dimension alignment, and the resulting visual prompt is fed into the U-Net cross-attention (Wu et al., 12 Sep 2025).

The stated rationale is twofold. First, image-derived prompts are more tightly coupled to the actual low-resolution input than generic or externally generated text descriptions, which can improve semantic consistency. Second, removing the text prompt extractor or VLM pipeline improves efficiency (Wu et al., 12 Sep 2025).

At inference time, RCOD supports two operating modes:

Mode	Mechanism	Reported role
Manual control	Select timestep $$\mathbf{z}_t = \alpha_t \mathbf{z}_0 + \sigma_t \boldsymbol{\epsilon}, \quad \alpha_t = \prod_{s=1}^t \sqrt{1-\beta_s}, \quad \sigma_t = \sqrt{1-\alpha_t^2}.$$2 directly	Choose fidelity, neutral, or realism
Adaptive control	Use the Metric Estimation Module (MEM)	Predict $$\mathbf{z}_t = \alpha_t \mathbf{z}_0 + \sigma_t \boldsymbol{\epsilon}, \quad \alpha_t = \prod_{s=1}^t \sqrt{1-\beta_s}, \quad \sigma_t = \sqrt{1-\alpha_t^2}.$$3 and map it to a timestep group

The Metric Estimation Module (MEM) is described as a simple MLP operating on intermediate features and trained after RCOD training using the same training data (Wu et al., 12 Sep 2025). Because $$\mathbf{z}_t = \alpha_t \mathbf{z}_0 + \sigma_t \boldsymbol{\epsilon}, \quad \alpha_t = \prod_{s=1}^t \sqrt{1-\beta_s}, \quad \sigma_t = \sqrt{1-\alpha_t^2}.$$4 is unavailable at test time, MEM provides an approximation to the latent degradation level needed for automatic group assignment. The paper reports that RCOD$$\mathbf{z}_t = \alpha_t \mathbf{z}_0 + \sigma_t \boldsymbol{\epsilon}, \quad \alpha_t = \prod_{s=1}^t \sqrt{1-\beta_s}, \quad \sigma_t = \sqrt{1-\alpha_t^2}.$$5-Adap. often lands in a middle group, yielding balanced performance (Wu et al., 12 Sep 2025).

The explicit inference interface in RCOD is therefore unusually simple for a controllable diffusion method: the timestep itself serves as the realism knob. A lower timestep preserves more structure from the input latent, while a higher timestep allows more generative synthesis of natural textures (Wu et al., 12 Sep 2025).

5. Empirical behavior, evaluation protocol, and efficiency

RCOD is trained on LSDIR, the first 10K FFHQ face images, and the Real-ESRGAN degradation pipeline (Wu et al., 12 Sep 2025). The evaluation uses DIV2K-Val with synthetic degradation, and the real-world datasets RealSR and DRealSR (Wu et al., 12 Sep 2025). The reported full-reference metrics are PSNR, SSIM, LPIPS, and DISTS, while the no-reference metrics are NIQE, MUSIQ, MANIQA, and CLIPIQA (Wu et al., 12 Sep 2025).

The framework is instantiated on two one-step backbones:

• OSEDiff $$\mathbf{z}_t = \alpha_t \mathbf{z}_0 + \sigma_t \boldsymbol{\epsilon}, \quad \alpha_t = \prod_{s=1}^t \sqrt{1-\beta_s}, \quad \sigma_t = \sqrt{1-\alpha_t^2}.$$6 RCOD$$\mathbf{z}_t = \alpha_t \mathbf{z}_0 + \sigma_t \boldsymbol{\epsilon}, \quad \alpha_t = \prod_{s=1}^t \sqrt{1-\beta_s}, \quad \sigma_t = \sqrt{1-\alpha_t^2}.$$7
• S3Diff $$\mathbf{z}_t = \alpha_t \mathbf{z}_0 + \sigma_t \boldsymbol{\epsilon}, \quad \alpha_t = \prod_{s=1}^t \sqrt{1-\beta_s}, \quad \sigma_t = \sqrt{1-\alpha_t^2}.$$8 RCOD$$\mathbf{z}_t = \alpha_t \mathbf{z}_0 + \sigma_t \boldsymbol{\epsilon}, \quad \alpha_t = \prod_{s=1}^t \sqrt{1-\beta_s}, \quad \sigma_t = \sqrt{1-\alpha_t^2}.$$9

The paper compares against multi-step diffusion methods (StableSR, ResShift, DiffBIR, SeeSR), one-step diffusion methods (SinSR, OSEDiff, S3Diff, TSD-SR, PiSA-SR, InvSR), and GAN-based methods (BSRGAN, Real-ESRGAN) (Wu et al., 12 Sep 2025).

Several concrete results are reported. For RCOD$$\hat{\mathbf{z}}_H = F_\theta(\mathbf{z}_L; t, c_y),$$0-Fid., the paper states that it obtains the best PSNR and SSIM among listed one-step methods on both DRealSR and RealSR, including DrealSR PSNR $$\hat{\mathbf{z}}_H = F_\theta(\mathbf{z}_L; t, c_y),$$1, SSIM $$\hat{\mathbf{z}}_H = F_\theta(\mathbf{z}_L; t, c_y),$$2 and RealSR PSNR $$\hat{\mathbf{z}}_H = F_\theta(\mathbf{z}_L; t, c_y),$$3, SSIM $$\hat{\mathbf{z}}_H = F_\theta(\mathbf{z}_L; t, c_y),$$4 (Wu et al., 12 Sep 2025). For RCOD$$\hat{\mathbf{z}}_H = F_\theta(\mathbf{z}_L; t, c_y),$$5-Real., the paper reports the lowest NIQE on RealSR, namely $$\hat{\mathbf{z}}_H = F_\theta(\mathbf{z}_L; t, c_y),$$6, together with very strong CLIPIQA and MUSIQ (Wu et al., 12 Sep 2025).

The evidence for realism control is the monotonic trend across the three timestep settings: Fid. ($$\hat{\mathbf{z}}_H = F_\theta(\mathbf{z}_L; t, c_y),$$7) improves PSNR, SSIM, and LPIPS; Neu. ($$\hat{\mathbf{z}}_H = F_\theta(\mathbf{z}_L; t, c_y),$$8) is intermediate; Real. ($$\hat{\mathbf{z}}_H = F_\theta(\mathbf{z}_L; t, c_y),$$9) improves NIQE, MUSIQ, and CLIPIQA (Wu et al., 12 Sep 2025). The paper also reports that RCOD preserves one-step efficiency, with A100 timings including RCOD$t=250$0-Fid.: $t=250$1, OSEDiff: $t=250$2, and PiSA-SR-adj.: $t=250$3 (Wu et al., 12 Sep 2025).

These findings support the claim that RCOD does not merely improve average quality; it exposes a controllable quality axis while maintaining the speed characteristic of one-step diffusion.

6. Relation to adjacent controllable one-step restoration methods

RCOD belongs to a broader family of one-step diffusion restoration methods that attempt to regulate the tension between generative richness and source faithfulness, but it does so through a distinctive timestep-centric design.

In CODSR, the controllable mechanism is distributed across three modules: LQ-guided feature modulation (LQFM) for fidelity, Region-adaptive generative prior activation (RGPA) for realism, and Text-matching guidance (TMG) for semantic alignment (Chen et al., 16 Dec 2025). CODSR identifies three limitations of prior one-step SR methods: information loss from VAE compression of low-quality inputs, insufficient region-discriminative activation of generative priors, and misalignment between text prompts and semantic regions (Chen et al., 16 Dec 2025). Its realism control is expressed through adaptive latent noise and timestep-dependent modulation, with the paper stating that the timestep $t=250$4 can be changed at inference to control the realism-fidelity tradeoff (Chen et al., 16 Dec 2025). Compared with RCOD, CODSR places greater emphasis on preserving uncompressed LQ information and grounding prompt semantics to spatial regions.

In HonestFace, the realism-control problem is specialized to face restoration rather than general Real-ISR. HonestFace uses a one-step latent diffusion restoration backbone combined with multi-reference conditioning, an identity embedder, masked face alignment (MFA), and an affine landmark distance metric (Wang et al., 24 May 2025). Its goal is “honest” face restoration that avoids over-smoothing, plastic-like skin, repetitive textures, color shift, and identity drift (Wang et al., 24 May 2025). Unlike RCOD, which uses the timestep as an explicit realism knob, HonestFace emphasizes identity-faithful conditioning and local perceptual supervision. The reported efficiency remains one-step, with Step: 1 and Inference time: $t=250$5 (Wang et al., 24 May 2025).

These comparisons clarify the scope of RCOD. RCOD is not simply “one-step diffusion plus extra losses.” Its defining feature is that it recovers a multi-regime fidelity–realism trade-off within a one-step model by structuring training and inference around grouped timesteps (Wu et al., 12 Sep 2025). CODSR and HonestFace are closely related in spirit, but they realize controllability through different mechanisms and problem formulations (Chen et al., 16 Dec 2025, Wang et al., 24 May 2025).

7. Practical significance, limitations, and interpretation

The practical significance of RCOD lies in the fact that real-world super-resolution does not admit a single universally optimal output style. The paper explicitly notes that some applications may prefer fidelity, such as restoration or documentation settings, whereas others may prefer realism, such as consumer photo enhancement (Wu et al., 12 Sep 2025). RCOD makes that preference a selectable inference-time parameter rather than an implicit property fixed by training.

Several limitations and caveats are also stated or implied in the reported discussion. First, the control mechanism is timestep-based and mainly discrete through group selection, rather than arbitrary continuous editing (Wu et al., 12 Sep 2025). Second, adaptive control depends on the Metric Estimation Module, whose prediction of degradation level is approximate (Wu et al., 12 Sep 2025). Third, replacing text guidance with VPIM may reduce some semantic priors; the paper notes that S3Diff’s negative online prompting is removed when using VPIM, which may help explain why DISTS is not always best (Wu et al., 12 Sep 2025). Fourth, the empirical validation is reported on two base OSD models, OSEDiff and S3Diff, which supports generality to a degree but does not amount to exhaustive backbone coverage (Wu et al., 12 Sep 2025).

Related work reinforces the same general trade-off. CODSR reports that increasing timestep increases realism metrics like MUSIQ but decreases PSNR, explicitly demonstrating a fidelity–quality trade-off (Chen et al., 16 Dec 2025). HonestFace shows that realism control in restoration can also depend on localized perceptual supervision and identity constraints, with MFA and adversarial distillation used to prevent over-smoothing and false textures (Wang et al., 24 May 2025). Taken together, these studies indicate that realism control in one-step diffusion restoration is not a single technique but an emerging design space.

In that design space, RCOD is distinguished by a precise and compact principle: control the denoising regime through degradation-conditioned timestep assignment. This makes the realism axis explicit, operational, and computationally compatible with one-step diffusion for real-world image super-resolution (Wu et al., 12 Sep 2025).