Faithful image super-resolution (SR) not only needs to recover images that appear realistic, similar to image generation tasks, but also requires that the restored images maintain fidelity and structural consistency with the input. To this end, we propose a simple and effective method, named FaithDiff, to fully harness the impressive power of latent diffusion models (LDMs) for faithful image SR. In contrast to existing diffusion-based SR methods that freeze the diffusion model pre-trained on high-quality images, we propose to unleash the diffusion prior to identify useful information and recover faithful structures. As there exists a significant gap between the features of degraded inputs and the noisy latent from the diffusion model, we then develop an effective alignment module to explore useful features from degraded inputs to align well with the diffusion process. Considering the indispensable roles and interplay of the encoder and diffusion model in LDMs, we jointly fine-tune them in a unified optimization framework, facilitating the encoder to extract useful features that coincide with diffusion process. Extensive experimental results demonstrate that FaithDiff outperforms state-of-the-art methods, providing high-quality and faithful SR results.

1. Bibliographic Information

• Title: FaithDiff: Unleashing Diffusion Priors for Faithful Image Super-resolution
• Authors: Junyang Chen, Jinshan Pan, Jiangxin Dong
• Affiliations: School of Computer Science and Engineering, Nanjing University of Science and Technology
• Journal/Conference: This paper is a preprint available on arXiv. It has not yet been published in a peer-reviewed journal or conference at the time of this analysis. arXiv is a common platform for researchers to share their work early.
• Publication Year: 2024
• Abstract: The paper introduces FaithDiff, a method for faithful image super-resolution (SR). Unlike existing methods that use frozen, pre-trained latent diffusion models (LDMs), FaithDiff proposes to "unleash" the diffusion prior by fine-tuning the model. This allows it to better identify useful structural information from degraded low-quality (LQ) inputs. The method includes a novel alignment module to bridge the feature gap between the LQ input and the noisy latent variables in the diffusion process. Crucially, FaithDiff jointly fine-tunes the image encoder and the diffusion model in a unified framework, enhancing their synergy. The authors claim that this approach leads to state-of-the-art results, producing SR images that are both high-quality and faithful to the original input.
• Original Source Link:
  • arXiv Page: https://arxiv.org/abs/2411.18824
  • PDF Link: http://arxiv.org/pdf/2411.18824v1

2. Executive Summary

• Background & Motivation (Why):

  • Core Problem: The central challenge in image super-resolution (SR) is to restore a high-quality (HQ) image from a low-quality (LQ) one. This is an ill-posed problem, meaning a single LQ image could correspond to many possible HQ images. The key is to achieve faithful SR, which requires the output to be not only realistic (like a generated image) but also structurally consistent with the original LQ input.
  • Gaps in Prior Work:
    1. GAN-based methods can produce sharp details but often introduce unrealistic artifacts and suffer from unstable training.
    2. Recent diffusion-based methods leverage powerful pre-trained latent diffusion models (LDMs) like Stable Diffusion. However, they typically freeze the diffusion model and only train an encoder to extract "degradation-robust" features from the LQ image to guide the generation. This approach has a critical flaw: if the encoder makes a mistake and extracts inaccurate features, the frozen diffusion model, trained only on pristine HQ images, can misinterpret these errors as actual image structures, leading to unfaithful results (e.g., garbled text, incorrect textures).
  • Innovation: FaithDiff challenges the "frozen prior" paradigm. Its core idea is that the diffusion model itself should be adapted to the task of restoration. By fine-tuning (unleashing) the diffusion model, it can learn to distinguish between genuine image content and degradation artifacts in the features provided by the encoder. This is complemented by jointly optimizing the encoder and the diffusion model, allowing them to co-adapt and work together more effectively.

• Main Contributions / Findings (What):

  1. Unleashing Diffusion Priors: The paper proposes to fine-tune the pre-trained LDM instead of keeping it frozen. This allows the model to better utilize information from degraded inputs and suppress reconstruction errors, leading to more faithful results.
  2. Effective Alignment Module: A simple yet effective module is introduced to align the features from the LQ input with the evolving noisy latent state of the diffusion model, ensuring that the guidance is relevant at each step of the denoising process.
  3. Unified Optimization Framework: The VAE encoder (which extracts LQ features) and the diffusion model are fine-tuned together. This synergy allows the encoder to learn to extract features that are most useful for the diffusion process, while the diffusion model learns to better interpret and refine these features.
  4. State-of-the-Art Performance: FaithDiff is shown to outperform existing methods on both synthetic and real-world benchmarks, particularly in perceptual quality and structural fidelity, as demonstrated by quantitative metrics and qualitative examples.

3. Prerequisite Knowledge & Related Work

Foundational Concepts

• Image Super-Resolution (SR): The task of increasing the resolution (i.e., pixel dimensions) of an image while maintaining or enhancing its quality. "Blind SR" refers to cases where the degradation process (e.g., blur, noise) is unknown.
• Generative Adversarial Networks (GANs): A class of deep learning models where two neural networks, a generator and a discriminator, compete against each other. The generator creates fake images, and the discriminator tries to distinguish them from real ones. In SR, the generator upscales the LQ image, aiming to fool the discriminator into thinking the result is a real HQ image. This can create sharp details but may also lead to artifacts.
• Diffusion Models (DDPMs): Generative models that learn to create data by reversing a gradual noising process.
  • Forward Process: Slowly add Gaussian noise to an image over many timesteps until it becomes pure noise.
  • Reverse Process: Train a neural network (typically a U-Net) to predict and remove the noise at each timestep, starting from random noise and gradually denoising it back into a clean image.
• Latent Diffusion Models (LDMs): A more efficient variant of diffusion models (e.g., Stable Diffusion). Instead of applying the diffusion process in the high-dimensional pixel space, LDMs first use a Variational Autoencoder (VAE) to compress the image into a smaller, lower-dimensional latent space. The diffusion process then happens in this latent space, which is computationally much cheaper. After denoising in the latent space, the VAE's decoder reconstructs the final HQ image.
• Variational Autoencoder (VAE): An architecture consisting of an encoder that maps an image to a latent representation and a decoder that reconstructs the image from that representation. LDMs use a pre-trained VAE.

Previous Works & Technological Evolution

The paper positions itself within the evolution of SR methods:

1. Early Methods: Focused on estimating the degradation kernel (e.g., blur) and then reversing it mathematically. These methods struggle with complex, real-world degradations.
2. GAN-based SR (e.g., Real-ESRGAN, BSRGAN): Introduced generative models to synthesize realistic textures. They significantly improved perceptual quality but were plagued by artifacts and training instability.
3. Diffusion-based SR (e.g., DiffBIR, PASD, SUPIR): The current state-of-the-art. These methods leverage the powerful generative priors of large pre-trained text-to-image LDMs. The common strategy is:
   • Keep the LDM (U-Net denoiser and VAE decoder) frozen.
   • Train a new encoder to extract features from the LQ input.
   • Use these features as a condition to guide the frozen LDM, often via mechanisms like ControlNet.
   • The core assumption is that the LDM's prior is perfect and only needs to be guided correctly.

Differentiation

FaithDiff's key departure from previous diffusion-based SR methods is its philosophy:

• Frozen Prior vs. Unleashed Prior: Instead of treating the LDM as a fixed, unchangeable prior, FaithDiff argues that the prior itself should be adapted to the SR task. Fine-tuning the LDM allows it to learn the statistics of restoration rather than just generation.

• Separate vs. Joint Optimization: Prior work often trains the guidance encoder separately. FaithDiff proposes a unified optimization framework where the feature encoder and the diffusion model are fine-tuned together. This creates a feedback loop: the encoder learns to produce features that the diffusion model can best use, and the diffusion model learns to better interpret the features the encoder provides. This synergy is a central claim of the paper.

  

As shown in Image 1, previous methods like DiffBIR and SUPIR rely on a degradation removal module (DRM) to produce features for a frozen diffusion model. If the DRM output is flawed (b, c), the final results (f, g) are unfaithful. FaithDiff (h) avoids this by jointly optimizing the components, leading to better structural recovery (e.g., the text).

4. Methodology (Core Technology & Implementation)

The FaithDiff pipeline, illustrated in Image 2, integrates several components into a unified, trainable framework.

Overall Pipeline:

1. An HQ image is encoded by a frozen VAE Encoder to get its latent representation $f^{HQ}$. Noise is added to this latent to create the noisy latent $x_t^{HQ}$ for a given timestep $t$. This is the training target.
2. The corresponding LQ image is fed into a trainable LQ Encoder (initialized from the VAE encoder) to extract LQ features $f^{LQ}$.
3. $f^{LQ}$ and the noisy latent $x_t^{HQ}$ are fed into the Alignment Module to produce aligned features $f_t^a$.
4. The aligned features $f_t^a$, along with optional text embeddings $c$, are used to condition the Diffusion Model (a U-Net), which predicts the noise $\hat{\epsilon}_\theta$ present in $x_t^{HQ}$.
5. The model is trained to minimize the difference between the predicted noise $\hat{\epsilon}_\theta$ and the actual added noise $\epsilon$.
6. During inference, the process starts with random noise and iteratively denoises it using the trained model, guided by the LQ image features, to produce a clean latent.
7. Finally, the frozen VAE Decoder converts the clean latent back into the restored HQ image.

4.1. LQ Feature Extraction

Instead of using the final layer of the VAE encoder, which is highly compressed (e.g., 8 channels), FaithDiff extracts features from the penultimate layer. This layer has a much higher channel dimension (512 channels), providing a richer representation that captures more detail about both the image structure and the degradation, which is beneficial for the subsequent diffusion process.

4.2. Alignment Module

There is a significant domain gap between the static LQ features $f^{LQ}$ and the noisy latent $x_t^{HQ}$, which becomes progressively cleaner as the diffusion process (denoising) proceeds. Simply adding them is suboptimal. The Alignment Module is designed to bridge this gap.

As shown in the bottom right of Image 2, the module works as follows:

1. The noisy latent $x_t^{HQ}$ and the LQ features $f^{LQ}$ are passed through separate $3 \times 3$ convolutional layers to get $f_t^x$ and $f^m$, respectively.

2. These two feature maps are concatenated: $f_t^c = \mathrm{Concat}(f_t^x, f^m)$.

3. The concatenated features are processed by a stack of two Transformer blocks, which use self-attention to allow for rich interaction between the information from the noisy latent and the LQ input.

4. A residual connection adds the original processed latent features $f_t^x$ back, and a final linear (fully connected) layer produces the aligned features $f_t^a$.

   The process is described by the following equations: $$\begin{array} { r l } & { f _ { t } ^ { x } = \mathrm { Conv } ( x _ { t } ^ { H Q } ) , f ^ { m } = \mathrm { Conv } ( f ^ { L Q } ) , } \\ & { f _ { t } ^ { c } = \mathrm { Concat } ( f _ { t } ^ { x } , f ^ { m } ) , } \\ & { \mathrm { Trans } ( f _ { t } ^ { c } ) = \mathcal { T } _ { 2 } ( \mathcal { T } _ { 1 } ( f _ { t } ^ { c } ) ) , } \\ & { f _ { t } ^ { a } = \mathrm { Linear } ( \mathrm { Trans } ( f _ { t } ^ { c } ) + f _ { t } ^ { x } ) , } \end{array}$$ where:

• $x_t^{HQ}$ is the noisy latent at timestep $t$.

• $f^{LQ}$ are the features from the LQ encoder.

• $\mathrm{Conv}(\cdot)$ is a $3 \times 3$ convolution operation.

• $\mathrm{Concat}(\cdot)$ is the concatenation operation.

• $\mathcal{T}_i$ denotes a Transformer block.

• $\mathrm{Linear}(\cdot)$ is a fully connected layer.

• $f_t^a$ are the final aligned features used to guide the diffusion model.

  The diffusion model then uses these aligned features $f_t^a$ and text embeddings $c$ to predict the noise, and one step of denoising is performed according to: $$x _ { t - 1 } ^ { H Q } = \frac { 1 } { \sqrt { \alpha _ { t } } } ( x _ { t } ^ { H Q } - \frac { 1 - \alpha _ { t } } { \sqrt { 1 - \bar { \alpha } _ { t } } } \hat { \epsilon } _ { \theta } ( f _ { t } ^ { a } , c , t ) ) + \sigma _ { t } z ,$$

• $\hat{\epsilon}_\theta$ is the diffusion model (U-Net) predicting the noise.

• $\alpha_t$, $\bar{\alpha}_t$, and $\sigma_t$ are standard noise schedule parameters from DDPMs.

• $z$ is random Gaussian noise.

4.3. Unified Feature Optimization

This is the core training strategy. The trainable components are the LQ Encoder, the Alignment Module, and the Diffusion Model. The VAE decoder and text encoder are kept frozen.

The training happens in two stages:

1. Pre-training the Alignment Module: Initially, only the alignment module is trained for 6,000 iterations. The LQ encoder and diffusion model are frozen. This helps the alignment module find a good initialization for connecting the two frozen components.

2. Joint Fine-tuning: All three components (LQ Encoder, Alignment Module, Diffusion Model) are trained together end-to-end for 40,000 iterations. This allows them to co-adapt, which is the key to achieving high fidelity.

   The optimization is driven by a simple L1 loss on the predicted noise: $$L = \left\| \epsilon - \hat { \epsilon } _ { \theta } ( \sqrt { \bar { \alpha } _ { t } x _ { 0 } ^ { H Q } } + \sqrt { 1 - \bar { \alpha } _ { t } } \epsilon , f ^ { L Q } , c , t ) \right\| _ { 1 }$$

• $L$ is the loss.
• $\epsilon$ is the ground-truth noise sampled from a standard normal distribution.
• $x_0^{HQ}$ is the latent representation of the clean HQ image.
• The term $\sqrt { \bar { \alpha } _ { t } x _ { 0 } ^ { H Q } } + \sqrt { 1 - \bar { \alpha } _ { t } } \epsilon$ is the noisy latent $x_t^{HQ}$.
• $\hat{\epsilon}_\theta(\cdot)$ is the noise predicted by the U-Net, conditioned on the noisy latent, the LQ features $f^{LQ}$, the text condition $c$, and the timestep $t$.
• $\|\cdot\|_1$ denotes the L1 norm (mean absolute error).

5. Experimental Setup

• Datasets:

  • Training: A large collection of high-resolution images from LSDIR, DIV2K, Flicker2K, DIV8K, and FFHQ (faces). LQ images are synthesized with a complex degradation model. Text descriptions are generated using LLAVA.
  • Synthetic Test: DIV2K-Val and LSDIR-Val with three different levels of degradation severity (Level-I, II, III).
  • Real-World Test: RealPhoto60 and a newly collected RealDeg dataset (238 images from old photos, films, etc.) to test generalization.

• Evaluation Metrics:

  • Fidelity Metrics (Reference-based):
    1. PSNR (Peak Signal-to-Noise Ratio): Measures the ratio between the maximum possible power of a signal and the power of corrupting noise. Higher is better. It focuses on pixel-wise differences. $$\mathrm{PSNR} = 10 \cdot \log_{10} \left( \frac{\mathrm{MAX}_I^2}{\mathrm{MSE}} \right)$$
       • $\mathrm{MAX}_I$: Maximum possible pixel value of the image (e.g., 255).
       • $\mathrm{MSE}$: Mean Squared Error between the ground truth and restored images.
    2. SSIM (Structural Similarity Index Measure): Measures the similarity between two images based on luminance, contrast, and structure. It aligns better with human perception of similarity than PSNR. A value of 1 indicates perfect similarity. Higher is better. $$\mathrm{SSIM}(x, y) = \frac{(2\mu_x\mu_y + c_1)(2\sigma_{xy} + c_2)}{(\mu_x^2 + \mu_y^2 + c_1)(\sigma_x^2 + \sigma_y^2 + c_2)}$$
       • $\mu_x, \mu_y$: Mean of images $x$ and $y$.
       • $\sigma_x^2, \sigma_y^2$: Variance of images $x$ and $y$.
       • $\sigma_{xy}$: Covariance of $x$ and $y$.
       • $c_1, c_2$: Small constants for stability.
  • Perceptual Quality Metrics:
    1. LPIPS (Learned Perceptual Image Patch Similarity): Measures the perceptual distance between two images using features from a deep neural network (e.g., VGG). Lower is better.
    2. MUSIQ (Multi-scale Image Quality Transformer): A no-reference metric that predicts image quality based on features from a Transformer model. Higher is better.
    3. CLIPIQA+: A no-reference metric that leverages the CLIP model to assess image quality by comparing the image to a range of quality-aware text prompts. Higher is better.
  • OCR Metrics:
    1. Precision: The fraction of detected text boxes that are correct.
    2. Recall: The fraction of ground-truth text boxes that are correctly detected.

• Baselines:

  • GAN-based: Real-ESRGAN, BSRGAN.
  • Diffusion-based: StableSR, DiffBIR, PASD, DreamClear, SeeSR, SUPIR.

6. Results & Analysis

Core Results

The paper presents extensive quantitative and qualitative comparisons.

• Synthetic Datasets (Table 1):

  The following table, transcribed from the paper, shows results on synthetic benchmarks.

  Table 1. Quantitative comparisons on synthetic benchmarks (DIV2K-Val and LSDIR-Val). Best and second best performances in perceptual-oriented metrics (LPIPS, MUSIQ, CLIPIQA+) are marked in red and blue, respectively.

  D-Level	Methods	DIV2K-Val [1]					LSDIR-Val [17]
  		PSNR (dB) ↑	SSIM ↑	LPIPS ↓	MUSIQ ↑	CLIPIQA+ ↑	PSNR (dB) ↑	SSIM ↑	LPIPS ↓	MUSIQ ↑	CLIPIQA+ ↑
  Level-I	Real-ESRGAN [40]	26.64	0.7737	0.1964	62.38	0.4649	23.47	0.7102	0.2008	69.23	0.5378
  	BSRGAN [46]	27.63	0.7897	0.2038	61.81	0.4588	24.42	0.7292	0.2167	66.21	0.5037
  	StableSR [38]	24.71	0.7131	0.2393	65.55	0.5156	21.57	0.6233	0.2509	70.52	0.6004
  	DiffBIR [21]	24.60	0.6595	0.2496	66.23	0.5407	21.75	0.5837	0.2677	68.96	0.5693
  	PASD [43]	25.31	0.6995	0.2370	64.57	0.4764	22.16	0.6105	0.2582	68.90	0.5221
  	SeeSR [41]	25.08	0.6967	0.2263	66.48	0.5336	22.68	0.6423	0.2262	70.94	0.5815
  	DreamClear [2]	23.76	0.6574	0.2259	66.15	0.5478	20.08	0.5493	0.2619	70.81	0.6182
  	SUPIR [44]	25.09	0.7010	0.2139	65.49	0.5202	21.58	0.5961	0.2521	71.10	0.6118
  	Ours	24.29	0.6668	0.2187	66.53	0.5432	21.20	0.5760	0.2264	71.25	0.6253
  Level-II	Real-ESRGAN [40]	25.49	0.7274	0.2309	61.84	0.4719	22.47	0.6567	0.2342	69.14	0.5456
  	BSRGAN [46]	26.42	0.7402	0.2465	60.00	0.4463	23.35	0.6682	0.2641	64.17	0.4858
  	StableSR [38]	24.26	0.6771	0.2590	64.76	0.5057	21.58	0.5946	0.2802	69.57	0.5667
  	DiffBIR [21]	24.42	0.6441	0.2708	64.83	0.5246	21.63	0.5672	0.2853	67.61	0.5555
  	PASD [43]	24.89	0.6764	0.2502	64.45	0.4718	21.85	0.5846	0.2737	68.53	0.5131
  	SeeSR [41]	24.65	0.6734	0.2428	66.09	0.5226	22.00	0.6026	0.2469	70.91	0.5837
  	DreamClear [2]	23.39	0.6330	0.2518	64.96	0.5295	19.74	0.5191	0.2910	70.41	0.6072
  	SUPIR [44]	24.42	0.6703	0.2432	65.58	0.5202	21.30	0.5713	0.2733	70.59	0.5998
  	Ours	23.80	0.6413	0.2407	66.42	0.5460	20.88	0.5493	0.2469	71.15	0.6219
  Level-III	Real-ESRGAN [40]	22.81	0.6288	0.3535	60.11	0.4637	20.13	0.5374	0.3650	67.02	0.5275
  	BSRGAN [46]	23.45	0.6281	0.3462	62.41	0.4838	20.75	0.5358	0.3667	67.41	0.5363
  	StableSR [38]	23.34	0.6277	0.3559	57.89	0.4124	20.55	0.5195	0.3716	64.31	0.4859
  	DiffBIR [21]	23.42	0.5992	0.3676	58.86	0.5154	20.53	0.4809	0.3951	62.23	0.5148
  	PASD [43]	22.58	0.5944	0.3646	63.08	0.4815	20.03	0.4974	0.3769	67.43	0.5092
  	SeeSR [41]	22.58	0.5985	0.3278	65.82	0.5106	20.16	0.5046	0.3437	69.35	0.5444
  	DreamClear [2]	21.82	0.5510	0.3336	62.59	0.4914	18.46	0.4341	0.3831	68.64	0.5757
  	SUPIR [44]	21.90	0.5611	0.3172	66.28	0.5275	19.17	0.4650	0.3488	70.16	0.5917
  	Ours	21.77	0.5662	0.3080	66.28	0.5275	18.92	0.4568	0.3170	71.37	0.6067

  • Analysis: FaithDiff does not achieve the highest PSNR or SSIM; GAN-based methods excel here as they are often optimized for these metrics. However, FaithDiff consistently achieves the best or second-best scores on the perceptual metrics (LPIPS, MUSIQ, $CLIPIQA+$). This indicates that its outputs are more visually pleasing and perceptually closer to the ground truth, which is the main goal of faithful SR. The advantage is particularly noticeable at Level-III (severe degradation), where FaithDiff's ability to handle ambiguity and restore structure shines.

• Qualitative Comparison (Figure 3):

  Figure 3. Image SR result $( \times 4 )$ on the synthetic benchmark. The restored image by GAN-based methods [40] exhibits perceptually in and Incnras he roethevmu leareas wi thulre ).

  This figure shows a heavily degraded image of a bird. Real-ESRGAN (b) produces unpleasant, artificial textures. Other diffusion methods like DiffBIR (d) and SUPIR (g) fail to restore the correct feather structure, introducing blur or incorrect patterns. In contrast, FaithDiff (h) recovers fine, realistic feather details that are consistent with the ground truth (e). This highlights the benefit of unleashing the diffusion prior to correctly interpret ambiguous LQ features.

• Real-World Datasets (Table 2 & Figure 4):

  The following table, transcribed from the paper, shows results on real-world benchmarks.

  Table 2. Quantitative comparisons on real-world benchmarks. Best and second best performances are highlighted in red and blue, respectively.

  Benchmarks	Metrics	Real-ESRGAN [40]	BSRGAN [46]	StableSR [38]	DiffBIR [21]	PASD [43]	SeeSR [41]	DreamClear [2]	SUPIR [44]	Ours
  RealPhoto60 [44]	MUSIQ ↑	59.29	45.46	57.89	63.67	64.53	70.80	70.46	70.26	72.74
  	CLIPIQA+	0.4389	0.3397	0.4214	0.4935	0.4786	0.5691	0.5273	0.5528	0.5932
  RealDeg	MUSIQ ↑	52.64	52.08	53.53	58.22	47.31	60.10	56.67	51.50	61.24
  	CLIPIQA+ ↑	0.3396	0.3520	0.3669	0.4258	0.3137	0.4315	0.4105	0.3468	0.4327

  FaithDiff achieves the best scores on both MUSIQ and $CLIPIQA+$ for both real-world datasets, demonstrating its superior generalization to diverse, authentic degradations.

  Figure 4. Image SR result $( \times 2 )$ on the real-world benchmarks. Compared to competing methods, our approach generates more realistic images with fine-scale structures and details.

  Figure 4 showcases results on real-world images. In all three examples (a flower, an old photo, a cityscape), FaithDiff (h) produces cleaner, more detailed, and more structurally correct results compared to the competitors.

Run-time and OCR Analysis

• Run-time (Table 4): FaithDiff is significantly faster than most diffusion-based competitors. It achieves high quality in just 20 inference steps and takes only 2.55 seconds for a 1024x1024 image. This efficiency comes from not needing heavy external adaptors like ControlNet.

  Table transcribed from the paper.

  Table 4. Run-time performance of diffusion-based SR methods for the diffusion process.

  	DiffBIR [21]	PASD [43]	SeeSR [41]	DreamClear [41]	SUPIR [44]	Ours
  Inference Step	50	20	50	50	50	20
  Running Time (s)	46.81	7.31	10.31	7.58	11.44	2.55

• OCR Recognition (Table 3): To objectively measure structural fidelity, the authors test OCR performance on restored images containing text. FaithDiff achieves the highest Precision (36.45%) and Recall (46.74%), significantly outperforming others. This provides strong evidence that it restores legible and accurate structures.

  Table transcribed from the paper.

  Table 3. OCR recognition performance on restored images.

  Metrics	GT	LQ	Real-ESRGAN [40]	BSRGAN [46]	StableSR [38]	DiffBIR [21]	PASD [43]	SeeSR [41]	DreamClear [2]	SUPIR [44]	Ours
  Precision	52.72%	7.54%	13.19%	12.04%	19.87%	26.21%	24.32%	30.07%	22.45%	31.78%	36.45%
  Recall	56.67%	7.59%	13.68%	12.33%	20.31%	27.90%	25.14%	33.09%	23.50%	41.57%	46.74%

Ablation Studies

• Effectiveness of Alignment Module (Table 5): The ablation study confirms the importance of each component of the alignment strategy. Removing the module (Ours w/o Align), skipping its pre-training (Ours w/o Pre-train align), or using features from the last layer instead of the penultimate layer (Ours w/ Last feats) all lead to significant performance degradation.

  Table transcribed from the paper.

  Table 5. Effectiveness of the proposed alignment module.

  	Alignment module	Pre-train alignment	Penultimate visual features	DIV2K-Val [1]	RealPhoto60 [44]
  				LPIPS↓	MUSIQ ↑
  Ours w/o Align	✗	✓	✓	0.3199	66.67
  Ours w/o Pre-train align	✓	✗	✓	0.3244	69.76
  Ours w/ Last feats	✓	✓	✗	0.3302	70.05
  Ours	✓	✓	✓	0.3080	72.74

• Effectiveness of Unified Feature Optimization (Table 6 & Figure 5): This is the most crucial ablation.

  • FT EN & Fix DM: Fine-tuning only the encoder while keeping the diffusion model fixed (the common approach) yields poor results.

  • Fix EN & FT DM: Fine-tuning only the diffusion model is better, but not optimal.

  • FT EN & DM (SP): Fine-tuning both, but separately, also underperforms.

  • Ours: Jointly fine-tuning both yields the best results, confirming the importance of their interplay.

    Table transcribed from the paper.

    Table 6. Effectiveness of unified feature optimization. SP denotes separately fine-tuning.

    	Encoder (EN)		Diffusion model (DM)		DIV2K-Val [1]	RealPhoto60 [44]
    	Fix	Fine-tune (FT)	Fix	Fine-tune (FT)	LPIPS↓	MUSIQ ↑
    FT EN & Fix DM	✗	✓	✓	✗	0.3370	69.66
    Fix EN & FT DM	✓	✗	✗	✓	0.3302	71.11
    FT EN & DM (SP)	✗	✓	✗	✓	0.3261	69.94
    Ours	✗	✓	✗	✓	0.3080	72.74

    Figure 5. Effectiveness of the unified feature optimization on image SR $( \times 4 )$ . Using unify optimization strategy is able to generate the results with clearer structural details.

    Figure 5 vividly illustrates this. When the diffusion model is fixed (FT EN & Fix DM), it misinterprets flawed LQ features and hallucinates "windows" where there should be the letters "MAER" (b). Unleashing the DM helps (c, e), but only the joint optimization in FaithDiff (f) correctly restores the text, proving the synergy is essential for fidelity.

  • Visualization of DAAMs (Figure 6, labeled as image 5):

    

    This visualization shows the Diffusion Attentive Attribution Maps (DAAMs) for the word "bottle". For a severely degraded LQ patch (a), competing methods (b, c) show weak or misplaced attention, leading to poor restoration (e, f). FaithDiff's approach results in a much stronger and more accurate attention map (d), allowing the LDM to correctly identify and restore the bottles (g). This shows that joint optimization helps the model "see" through the degradation more effectively.

  7. Conclusion & Reflections

  • Conclusion Summary: The paper successfully argues that for faithful image super-resolution, the common practice of using a frozen diffusion prior is suboptimal. By "unleashing" the prior through fine-tuning and jointly optimizing the feature encoder and the diffusion model, FaithDiff creates a more powerful and synergistic system. This approach allows the model to better distinguish content from degradation, leading to state-of-the-art results in both perceptual quality and structural fidelity. The proposed alignment module and efficient design further contribute to its effectiveness.

  • Limitations & Future Work:

    • Computational Cost: Although faster at inference, fine-tuning a large LDM like SDXL is computationally expensive and requires significant GPU resources, which may be a barrier for some researchers.
    • Data Dependency: The model's performance on unseen degradation types is still dependent on the diversity of the training data. While it generalizes well, extreme or unusual artifacts might still pose a challenge.
    • Text Prompt Reliance: The method can leverage text prompts, but the quality of restoration might depend on the accuracy and relevance of these prompts. The paper uses an automatic captioning model (LLAVA), which is not perfect.

  • Personal Insights & Critique:

    • Paradigm Shift: The core idea of fine-tuning the diffusion model itself is a simple but profound shift from the prevailing trend in the field. It moves away from treating large models as black-box feature extractors and towards adapting them holistically for specific downstream tasks. This philosophy could be highly influential for other image restoration problems like deblurring, inpainting, and colorization.
    • Simplicity and Effectiveness: The proposed alignment module is straightforward (convs + transformers), yet the results show it is highly effective. This demonstrates that clever architectural design and a sound training strategy can be more impactful than overly complex modules.
    • Future Impact: FaithDiff sets a new standard for diffusion-based SR. Future work will likely build on this "unleashed prior" concept. We might see explorations into more parameter-efficient fine-tuning techniques (like LoRA) to reduce the computational burden, or methods that can adapt the diffusion prior on-the-fly for a specific input image. The emphasis on joint optimization of all components is a valuable lesson for designing future deep learning systems for image restoration.