QUSR: Quality-Aware & Uncertainty-Guided SR

• QUSR is a diffusion-based image super-resolution framework that uses quality-aware textual priors and uncertainty-guided noise modulation to address spatially non-uniform degradations.
• It employs a one-step residual diffusion model in latent space by combining a Quality-Aware Prior from Qwen2.5-VL-7B with Uncertainty-Guided Noise Generation for adaptive restoration.
• Empirical results on RealSR and DRealSR benchmarks demonstrate improved perceptual quality and fine detail reconstruction, balancing global semantic guidance with local fidelity.

Searching arXiv for QUSR and the related image super-resolution papers mentioned in the provided data. First, I’ll look up the main QUSR paper by title and then fetch IDs for key related work such as StableSR, DiffBIR, SeeSR, PiSA-SR, XPSR, SinSR, and OSEDiff. Searching "QUSR Quality-Aware and Uncertainty-Guided Image Super-Resolution Diffusion Model" QUSR, short for Quality-Aware and Uncertainty-Guided Super-Resolution, is a diffusion-based image super-resolution framework designed for real-world, unknown, and spatially non-uniform degradations. It addresses a central difficulty in practical ISR: some regions of a low-quality image remain relatively trustworthy, while others are heavily corrupted by blur, noise, compression, motion, or other complex degradations. QUSR combines a Quality-Aware Prior (QAP) derived from Qwen2.5-VL-7B with an Uncertainty-Guided Noise Generation (UNG) mechanism that modulates diffusion noise spatially, aiming to preserve reliable information in low-uncertainty regions while enabling aggressive detail synthesis in high-uncertainty regions (Yin et al., 10 Mar 2026).

1. Problem formulation and research context

In the formulation used by QUSR, ISR reconstructs a high-quality image $x_{hq}$ from a degraded low-quality image $x_{lq}$. The paper distinguishes this setting from controlled super-resolution benchmarks based on known bicubic down-sampling: in real-world conditions, the degradation pipeline is unknown, complex, non-linear, and often spatially non-uniform. Some regions may be sharp, while others are severely blurred or noisy, making the inverse problem highly ill-posed (Yin et al., 10 Mar 2026).

The QUSR paper situates itself against several diffusion-based and prompt-conditioned ISR approaches, including StableSR, DiffBIR, SeeSR, PiSA-SR, XPSR, and UPSR, and also aligns its one-step residual design with SinSR and OSEDiff (Yin et al., 10 Mar 2026). Its critique of prior work is organized around two persistent issues. The first is the tension between global priors and local fidelity: high-level semantic guidance can improve realism, but it may overlook low-level degradation and spatially varying reconstruction difficulty; conversely, direct conditioning on low-quality image features can be unreliable because those features are already corrupted. The second issue is spatially uniform noise in diffusion: applying the same noising intensity everywhere can destroy useful information in flat regions while failing to stimulate sufficient detail synthesis in edges and textures (Yin et al., 10 Mar 2026).

The core claim of QUSR is that these two issues should be handled jointly. Its “dual guidance” consists of a global quality-aware textual prior and a local uncertainty-aware heteroscedastic noising scheme, so that semantic and degradation descriptions steer the restoration globally while spatial uncertainty modulates the latent perturbation and training loss locally (Yin et al., 10 Mar 2026).

2. Backbone architecture and one-step residual diffusion

QUSR is built on Stable Diffusion v2.1 in latent space and adopts a single-step residual diffusion design rather than a multi-step denoising chain. The low-quality image is first encoded by a VAE encoder $\mathcal{E}$:

$\boldsymbol{z_{lq}} = \mathcal{E}(x_{lq}).$

UNG then perturbs this latent representation to produce a guided latent:

$$\boldsymbol{z_g} = \text{UNG}(\boldsymbol{z_{lq}}, x_{lq}).$$

A UNet denoising model $f(\cdot)$, fine-tuned with LoRA on top of the Stable Diffusion UNet, predicts a residual noise term at fixed timestep $t=1$ under quality-aware conditioning $\boldsymbol{C}_q$:

$$\boldsymbol{\epsilon_g} = f(\boldsymbol{z_g}, t, \boldsymbol{C}_q).$$

The restored latent is then formed by subtracting the predicted residual:

$$\boldsymbol{z_{hq}} = \boldsymbol{z_{lq}} - \boldsymbol{\epsilon_g},$$

and the VAE decoder $x_{lq}$0 maps it back to image space:

$x_{lq}$1

This formulation makes QUSR a one-step residual diffusion model in latent space, with the global “time” parameter fixed to $x_{lq}$2 during both training and inference (Yin et al., 10 Mar 2026).

This architectural choice is significant because it departs from the standard diffusion interpretation of a long forward–reverse chain. In QUSR, the forward perturbation is explicitly constructed and spatially modulated by uncertainty, and the reverse process is collapsed into a single UNet pass. A common simplification is to describe the method as merely “Stable Diffusion plus text prompts,” but that omits two essential design decisions: the model is residual, not iterative, and its latent perturbation is heteroscedastic, not spatially uniform (Yin et al., 10 Mar 2026).

3. Quality-Aware Prior (QAP)

The Quality-Aware Prior is a textual prior that jointly describes semantic content and quality/degradation attributes. The paper contrasts this with content-only prompting: instead of only describing what appears in the image, QAP is intended to encode clarity, noise, color, and lighting as well (Yin et al., 10 Mar 2026).

The low-quality image $x_{lq}$3 is fed into Qwen2.5-VL-7B-Instruct with the prompt:

“Please describe this low-resolution image, evaluating its quality based on clarity, color, noise, and lighting.”

The resulting text $x_{lq}$4 is encoded by a CLIP text encoder:

$x_{lq}$5

These embeddings condition the diffusion UNet through cross-attention at every layer. For the $x_{lq}$6-th feature map $x_{lq}$7, the conditioned feature is written as:

$x_{lq}$8

where $x_{lq}$9, $\mathcal{E}$0, and $\mathcal{E}$1 are learned linear projections and $\mathcal{E}$2 is the scaling factor (Yin et al., 10 Mar 2026).

Within the QUSR framework, QAP serves two roles. First, it supplies global semantic understanding and explicit quality-aware guidance to the denoiser. Second, it provides a degree of interpretability, because the textual description exposes the model’s high-level assessment of degradation. The paper emphasizes that Qwen’s judgments match human perception in the examples shown in Figure 1(a), and it argues that this makes restoration decisions more inspectable (Yin et al., 10 Mar 2026).

QAP also participates in the training objective through Classifier Score Distillation (CSD). The paper describes a pre-trained Stable Diffusion model as an “implicit classifier” used with classifier-free guidance to encourage generated images to remain consistent with the quality-aware prompt. The exact formula is not fully specified in the paper, but the role of $\mathcal{E}$3 is to align generation with the semantic and quality prior encoded in $\mathcal{E}$4 (Yin et al., 10 Mar 2026).

4. Uncertainty-Guided Noise Generation (UNG)

The second defining component of QUSR is Uncertainty-Guided Noise Generation, which introduces spatially varying noise into the latent representation. Its motivation is explicit: flat regions with reliable observations should not be aggressively noised, whereas edges, textures, and other ambiguous regions may require stronger perturbation to enable plausible detail reconstruction (Yin et al., 10 Mar 2026).

UNG begins with an Uncertainty Estimation Module (UEM), implemented as a lightweight encoder–decoder CNN. The encoder consists of 3 × $\mathcal{E}$5 conv layers with ELU activations, and the decoder mirrors this with 3 × $\mathcal{E}$6 conv layers and no final activation. Given the low-quality image, the UEM predicts a raw uncertainty map:

$\mathcal{E}$7

where $\mathcal{E}$8 is interpreted as a pixelwise aleatoric error scale (Yin et al., 10 Mar 2026).

The uncertainty map is then projected into latent space and transformed into a spatially varying noise standard deviation. The sequence is:

$\mathcal{E}$9

$\boldsymbol{z_{lq}} = \mathcal{E}(x_{lq}).$0

$\boldsymbol{z_{lq}} = \mathcal{E}(x_{lq}).$1

where $\boldsymbol{z_{lq}} = \mathcal{E}(x_{lq}).$2 is a scaling factor, $\boldsymbol{z_{lq}} = \mathcal{E}(x_{lq}).$3 enforces a minimum uncertainty, and $\boldsymbol{z_{lq}} = \mathcal{E}(x_{lq}).$4 stabilizes the square root (Yin et al., 10 Mar 2026). The guided latent is then obtained by

$\boldsymbol{z_{lq}} = \mathcal{E}(x_{lq}).$5

with $\boldsymbol{z_{lq}} = \mathcal{E}(x_{lq}).$6 and scalar perturbation strength $\boldsymbol{z_{lq}} = \mathcal{E}(x_{lq}).$7 (Yin et al., 10 Mar 2026).

The training objective mirrors this heteroscedastic interpretation. QUSR includes an uncertainty loss

$\boldsymbol{z_{lq}} = \mathcal{E}(x_{lq}).$8

where the factor $\boldsymbol{z_{lq}} = \mathcal{E}(x_{lq}).$9 relaxes reconstruction constraints in high-uncertainty regions and strengthens them in low-uncertainty regions, while the regularizer prevents trivial inflation of uncertainty everywhere (Yin et al., 10 Mar 2026).

A recurring misconception would be to treat UNG as only an auxiliary confidence map. In QUSR it is more consequential than that: uncertainty affects both the latent perturbation used by the one-step diffusion process and the heteroscedastic reconstruction weighting used during training (Yin et al., 10 Mar 2026).

5. Optimization, data, and quantitative results

The full objective is a weighted sum of four terms:

$$\boldsymbol{z_g} = \text{UNG}(\boldsymbol{z_{lq}}, x_{lq}).$$0

with $$\boldsymbol{z_g} = \text{UNG}(\boldsymbol{z_{lq}}, x_{lq}).$$1, $$\boldsymbol{z_g} = \text{UNG}(\boldsymbol{z_{lq}}, x_{lq}).$$2, $$\boldsymbol{z_g} = \text{UNG}(\boldsymbol{z_{lq}}, x_{lq}).$$3, and $$\boldsymbol{z_g} = \text{UNG}(\boldsymbol{z_{lq}}, x_{lq}).$$4 (Yin et al., 10 Mar 2026). Here $$\boldsymbol{z_g} = \text{UNG}(\boldsymbol{z_{lq}}, x_{lq}).$$5 is the pixel-wise $$\boldsymbol{z_g} = \text{UNG}(\boldsymbol{z_{lq}}, x_{lq}).$$6 loss,

$$\boldsymbol{z_g} = \text{UNG}(\boldsymbol{z_{lq}}, x_{lq}).$$7

$$\boldsymbol{z_g} = \text{UNG}(\boldsymbol{z_{lq}}, x_{lq}).$$8 is the LPIPS perceptual loss, $$\boldsymbol{z_g} = \text{UNG}(\boldsymbol{z_{lq}}, x_{lq}).$$9 is the classifier score distillation term, and $f(\cdot)$0 is the uncertainty-aware term (Yin et al., 10 Mar 2026).

The implementation details are specific. QUSR uses Stable Diffusion 2.1 latent UNet + VAE, applies LoRA with rank 4 on the UNet, takes $f(\cdot)$1 low-quality inputs, and produces $f(\cdot)$2 outputs at $f(\cdot)$3 scale. Training uses LSDIR + the first 10k FFHQ images, with low-quality/high-quality pairs synthesized using the RealESRGAN degradation pipeline. Testing uses RealSR and DRealSR, with center-cropped pairs in the same $f(\cdot)$4 setup. Optimization is performed on 4 × NVIDIA RTX 3090 (24 GB) with Adam, learning rate $f(\cdot)$5, batch size 4, and 15k iterations. Qwen2.5-VL-7B-Instruct is used offline to generate prompts, which are then encoded by CLIP (Yin et al., 10 Mar 2026).

The evaluation protocol combines reference-based metrics—PSNR, SSIM, LPIPS, DISTS, and FID—with no-reference image quality assessment metrics—CLIPIQA, MUSIQ, and MANIQA (Yin et al., 10 Mar 2026).

Benchmark	QUSR results	Reported interpretation
RealSR	PSNR 25.54; SSIM 0.7289; LPIPS 0.2974; FID 125.27; CLIPIQA 0.6824; MANIQA 0.6564; MUSIQ 69.17	2nd best PSNR, best CLIPIQA and MANIQA, FID very close to OSEDiff’s 123.50, MUSIQ close to PiSA-SR’s 70.15
DRealSR	PSNR 29.81; SSIM 0.8200; LPIPS 0.2708; FID 113.87; CLIPIQA 0.7082; MUSIQ 67.00; MANIQA 0.6415	SOTA across all metrics; compared with second-best PiSA-SR, FID improves from 130.61 to 113.87 and MUSIQ from 66.11 to 67.00

These results support the paper’s main empirical claim: QUSR is especially effective under genuinely real-world degradations, where quality-aware prompting and uncertainty-aware perturbation are most consequential (Yin et al., 10 Mar 2026). The qualitative analysis in Figure 2 is consistent with this interpretation, reporting sharper and more structurally accurate details in complex patterns and fewer artifacts in flat regions (Yin et al., 10 Mar 2026).

6. Ablations, trade-offs, and limitations

The ablation study on DRealSR isolates the contributions of QAP and UNG. The baseline without QAP or UNG achieves PSNR 29.05, SSIM 0.8071, CLIPIQA 0.6745, MUSIQ 65.54, and MANIQA 0.6268. The w/o QAP variant, which retains UNG only, reaches PSNR 30.19, SSIM 0.8206, CLIPIQA 0.6853, MUSIQ 66.63, and MANIQA 0.6318. The w/o UNG variant, which retains QAP only, yields PSNR 29.74, SSIM 0.8179, CLIPIQA 0.6906, MUSIQ 66.58, and MANIQA 0.6392. The full QUSR model gives PSNR 29.81, SSIM 0.8200, CLIPIQA 0.7082, MUSIQ 67.00, and MANIQA 0.6415 (Yin et al., 10 Mar 2026).

These numbers establish a specific trade-off. Removing QAP slightly increases PSNR and SSIM, but it harms CLIPIQA, MUSIQ, and MANIQA. The paper interprets this as evidence that QAP pushes the model toward more perceptually natural, though not always maximally pixel-exact, reconstructions. By contrast, adding UNG on top of QAP improves all reported metrics relative to the w/o UNG variant, particularly the no-reference IQA metrics, indicating better fine-grained texture reconstruction and reduced oversmoothing (Yin et al., 10 Mar 2026).

The behavioral analysis further reports that the uncertainty maps highlight edges and complex textures as high-uncertainty regions and flat/background regions as low-uncertainty regions, consistent with the intended design. The paper also states that performance depends on the quality of the MLLM-generated descriptions, but argues that QUSR is more robust than a dual-prompt dependency such as XPSR because QAP is combined with low-quality image features and UNG rather than used in isolation (Yin et al., 10 Mar 2026).

The paper’s explicit limitations are brief, but the methodology suggests several likely constraints. Computational cost remains nontrivial because inference requires VAE encoding/decoding, a UNet pass, and evaluation of the UEM, and prompt generation with Qwen2.5-VL, although offline, is heavier than a small prompt encoder. Dependence on MLLM quality introduces possible failure modes if Qwen misjudges blur or noise. Extreme degradations and domain shift may still lead to hallucinations or artifacts, as is typical in generative restoration. A plausible implication is that the one-step residual formulation trades some flexibility for efficiency relative to full multi-step diffusion (Yin et al., 10 Mar 2026).

Within the broader ISR literature, QUSR is best understood not as a generic text-conditioned super-resolution model, but as a framework that explicitly couples quality-aware semantic conditioning with uncertainty-aware heteroscedastic diffusion behavior. That coupling is the paper’s central contribution, and it is the mechanism through which QUSR targets the longstanding tension between perceptual realism and local fidelity in real-world super-resolution (Yin et al., 10 Mar 2026).