HiFiPix: Bringing High-Fidelity Real-World Super-Resolution Back to Pixel Space

Wenjie Li; Du Chen; Hui Wang; Linlong Fan; Youwei Lv; Zeyi Chen; Zehan Chen; Jinwei Chen

cs.CVOct 26, 2026

HiFiPix: Bringing High-Fidelity Real-World Super-Resolution Back to Pixel Space

Authors: Wenjie Li, Du Chen, Hui Wang, Linlong Fan, Youwei Lv, Zeyi Chen, Zehan Chen, Jinwei Chen,

Organizations: Beijing University of Posts and Telecommunications · vivo BlueImage Lab · Zhejiang University

Abstract

Diffusion-based real-world image super-resolution suffers from limited reconstruction fidelity due to information loss in VAE-based latent representations. Recently, pixel-space diffusion models offer a promising high-faithfulness alternative by avoiding latent compression, but adapting them to Real-ISR remains challenging due to inherent patch boundary artifacts and degraded generative capability. To address the issues and maximize the potential of pretrained pixel-space diffusion models, we propose HiFiPix, a pixel-space Real-ISR framework capable of ultra high-fidelity reconstruction and fine-grained visual generation effects. To suppress patch boundary artifacts, we penalize unrealistic high-frequency periodic components with a Cross-Scale Wavelet Discriminator (CWSD) for image-level supervision, and Mixup Contrastive Self-Distillation (MixCSD) for intermediate representation regularization. Furthermore, we uncover a heterogeneous generation-fidelity trade-off across low-rank adapters (LoRA) and propose Selective LoRA Pruning (LoRA-SP) to adaptively prune LoRA layers, enhancing generative capability with minimal deterioration of reconstruction fidelity. Extensive experiments demonstrate that HiFiPix consistently outperforms existing state-of-the-art latent space methods in terms of both restoration accuracy and perceptual quality by a large margin. Code and models will be released.

Figures & tables

Figure 1: We propose HiFiPix, a Real-ISR framework built on a pixel-space large generative model, achieving high-fidelity restoration of real-world LQ images and superior perceptual quality compared with existing latent-space Real-ISR methods.

Figure 2: Fine-tuning pixel models faces two challenges compared with latent models: (a) without VAE decoders to preserve spatial continuity, pixel-space models are prone to patch boundary artifacts; (b) training progress tends to favor high-fidelity restoration at the expense of generative capability.

Figure 3: (a) Heavier boundary artifacts when super-resolving extremely blurry input. (b) Removal of individual LoRA layers reveals trade-offs between reconstruction and generative capabilities.

Figure 4: Architecture of HiFiPix. We introduce MixCSD and CSWD to alleviate patch boundary artifacts, and propose LoRA-SP to enhance the model’s generative capability.

Datasets	Metrics	SUPIR -s50	FaithDiff -s20	Dit4SR -s40	VOSR -s25	LucidFlux -s15	TSDSR -s1	PiSA-SR -s1	TVT -s1	ODTSR -s1	StrSR -s1	Ours -s1
DrealSR	PSNR $\uparrow$	26.87	27.29	25.79	27.72	20.82	27.71	28.31	28.25	28.15	23.90	29.62
	SSIM $\uparrow$	0.6702	0.7104	0.6817	0.7376	0.5806	0.7543	0.7804	0.7898	0.7740	0.6948	0.7990
	LPIPS $\downarrow$	0.4339	0.3497	0.3656	0.3664	0.4532	0.2997	0.2960	0.2900	0.2591	0.3345	0.2898
	DISTS $\downarrow$	0.2793	0.2386	0.2437	0.2446	0.2771	0.2153	0.2169	0.2206	0.1926	0.2377	0.2302
	MANIQA $\uparrow$	0.5281	0.6378	0.6232	0.5204	0.251	0.5883	0.6161	0.5784	0.6226	0.5598	0.6429
	MUSIQ $\uparrow$	53.91	65.79	64.51	57.70	67.97	66.83	66.11	65.60	62.83	60.98	68.18

Table 1: Quantitative results on real-world benchmarks, with the best/second-best results highlighted in bold / underlined . "-s" denotes the number of diffusion steps.

Methods	SUPIR -s50	FaithDiff -s20	Dit4SR -s40	VOSR-1.4B -s25	LucidFlux -s15	TSDSR -s1	PiSA-SR -s1	TVT -s1	ODTSR -s1	StrSR -s1	Ours -s1
Inference Time (s) $\downarrow$	10.8	5.0	57.6	7.9	23.8	0.48	0.45	2.4	5.2	0.97	0.27
GPU Memory (G) $\downarrow$	52.46	11.10	22.06	6.91	35.0	6.99	6.89	14.62	40.83	27.69	17.63

Table 2: Comparison of computational costs on GGT-GeneralMix test set.

Figure 5: Qualitative comparisons with existing Real-ISR methods on real-world test sets.

Table 9

Table 10

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	# of	Resolution	MANIQA $\uparrow$	MUSIQ $\uparrow$	Real	Real	Construction
Benchmark	Pairs	Resolution	of GT	of GT	LQ	GT	Manner
DIV2K	100	$512^{2}$	0.6274	63.7	✗	✓	Synthesized LQ
RealSR	100	$512^{2}$	0.6017	57.5	✓	✓	DSLR acquisition
DrealSR	100	$512^{2}$	0.5589	50.1	✓	✓	DSLR acquisition
ScreenSR	130	$512^{2}$	0.7187	72.2	✓	✓	Screen photography
GGT-GeneralMix	250	$1024^{2}$	0.6488	68.3	✓	✗	Generated GT

Appendix

Table 8: Comparisons between existing paired Real-ISR benchmarks and our Dual-Q-Bench.

Figure 7: (a) Insufficient GT Quality. Existing Real-ISR benchmarks often provide GT with limited sharpness. Consequently, restored results may contain substantially richer or sharper details than GT, achieving higher MUSIQ and MANIQA scores than GT. This mismatch makes such GTs unreliable for jointly evaluating restoration fidelity and visual generation effects. (b) Background De-focus in GT. Existing benchmarks, such as ScreenSR, contain GT images with de-focused backgrounds. Real-ISR models often recover details in these regions that are absent from the GT, increasing the discrepancy with the reference and consequently challenging fidelity evaluation.

Datasets	Metrics	SUPIR -s50	FaithDiff -s20	Dit4SR -s40	VOSR -s25	LucidFlux -s15	TSDSR -s1	PiSA-SR -s1	TVT -s1	ODTSR -s1	StrSR -s1	Ours -s1
RealSR	PSNR $\uparrow$	25.09	25.28	23.39	25.29	18.58	24.80	25.50	25.81	25.07	21.66	26.70
	SSIM $\uparrow$	0.6728	0.7073	0.6661	0.7150	0.5544	0.7159	0.7418	0.7597	0.7361	0.6531	0.7553
	LPIPS $\downarrow$	0.3731	0.2872	0.3193	0.2961	0.3856	0.2997	0.2672	0.2593	0.2395	0.3126	0.2622
	DISTS $\downarrow$	0.2515	0.2110	0.2234	0.2226	0.2658	0.2153	0.2044	0.2061	0.1892	0.2358	0.2191
	MANIQA $\uparrow$	0.5632	0.6737	0.6569	0.6510	0.6645	0.6342	0.6551	0.6225	0.6620	0.5967	0.6739
	MUSIQ $\uparrow$	57.83	68.91	67.72	70.47	71.40	71.23	70.15	69.86	68.29	66.15	70.78

Appendix

Table 9: Quantitative comparisons with existing Real-ISR methods on real-world benchmarks, with the best/second-best results in bold / underlined . "-s" denotes the number of diffusion steps.

Methods	SUPIR -s50	FaithDiff -s20	Dit4SR -s40	VOSR -s25	LucidFlux -s15	TSDSR -s1	PiSA-SR -s1	TVT -s1	ODTSR -s1	StrSR -s1	Ours -s1
Generative Model params (B) $\downarrow$	4.80	3.51	8.38	1.74	14.03	7.79	1.29	1.26	21.86	6.24	7.77
Language Model Params (B) $\downarrow$	13.35	13.35	13.35	0	0	0	0	0.47	8.29	4.44	0
Trainable Parts Params (M) $\downarrow$	1243.4	2612.2	350.9	1439	1630.8	95.2	8.1	30.07	1231.2	641.6	207.1

Appendix

Table 10: Comparison of model parameters, where the generative model includes the text encoder, the VAE encoder/decoder, and the diffusion backbones; a language model is utilized for generating image captions. Since HiFiPix does not require image captions or text encoders, the final model parameters are fewer than the original

9B

HiDream-O1-Image (with text encoder).

Layers	PSNR $\uparrow$	LPIPS $\downarrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$
1 layer	26,11	0.2764	70.15	0.6207
2 layers	26.09	0.2768	70.51	0.6236
4 layers (Ours)	26.06	0.2775	70.92	0.6274
6 layers	25.92	0.2798	71.07	0.6282
8 layers	25.63	0.2856	71.22	0.6298

Appendix

Table 11: Ablation on the number of selected layers in our LoRA-SP.

Figure 8: Qualitative comparisons with existing Real-ISR methods on paired LQ-GT real-world test sets: RealSR and DrealSR.

Figure 9: Qualitative comparisons with existing methods on unpaired RealLQ250 and RealLR200 test sets.

Figure 10: Qualitative comparisons with existing Real-ISR methods on paired LQ-GT real-world test sets: GGT-GeneralMix.

Figure 11: Qualitative comparisons with existing Real-ISR methods on our paired LQ-GT real-world benchmark: Dual-Q-Bench.

Image Super-Resolution

Image super-resolution reconstructs high-resolution images from lower-resolution observations, balancing detail, fidelity, and computational cost.

See all 250 papers

VICR: Visual In-Context Restoration for Real-World Image Super-Resolution
May 30, 2026Qichang Zhang, Hailong Wang, Baiang Li +4Diffusion Models for Image Restoration Image Super-Resolution
Faculty of Science and Technology, University of Macau · 2Nullmax · 3Hefei University of Technology +1
PixelIR: Fidelity-Perception Decoupling via Pixel-Space Image-Residual Flow Matching for Efficient One-Step Real-World Super-Resolution
Aug 31, 2026Bingtian Qiao, Yue Shi, Yong Guo +2Image Super-Resolution
Shanghai Jiao Tong University · Fuzhou University · Shanghai AI Laboratory
PixelSR: Pixel-Space Orthogonal Flow Decoupling for Real-World Image Super-Resolution
Oct 20, 2026Xiang Lv, Kai Zhao, Qiang Wang +3Image Super-Resolution Pixel-Space Diffusion Models
China University of Petroleum (East China) · Kuaishou Technology · Shenzhen University of Advanced Technology

May 30, 2026J/K move · Enter open · S save

VICR: Visual In-Context Restoration for Real-World Image Super-Resolution

Qichang Zhang, Hailong Wang, Baiang Li, Linhao Wang+3

Faculty of Science and Technology, University of Macau · 2Nullmax · 3Hefei University of Technology+1

Diffusion Models for Image Restoration Image Super-Resolution

Open paper AskPDF arXiv

Real-world image super-resolution (Real-ISR) requires balancing structural fidelity to degraded observations with realistic detail synthesis. However, existing generative Real-ISR methods often rely on entangled conditioning mechanisms, leading to structural drift or semantically inconsistent details. To address this issue, we propose Visual In-Context Restoration (VICR), a Diffusion Transformer (DiT)-based framework that formulates Real-ISR as image completion. Specifically, we introduce a decoupled visual prior injection mechanism that derives local and global cues from the low-quality (LQ) image: local cues help recover image structures and support high-frequency detail synthesis, while global cues guide overall generation and promote semantic consistency. For ambiguous regions under severe degradation, VICR employs an inference-time agent to refine semantic prompts using visual evidence from the LQ input while keeping model parameters fixed. Experiments show that VICR achieves state-of-the-art performance across multiple Real-ISR benchmarks with only 127M trainable parameters.

Abstract

Image Super-Resolution

Explore similar work

VICR: Visual In-Context Restoration for Real-World Image Super-Resolution

PixelIR: Fidelity-Perception Decoupling via Pixel-Space Image-Residual Flow Matching for Efficient One-Step Real-World Super-Resolution

PixelSR: Pixel-Space Orthogonal Flow Decoupling for Real-World Image Super-Resolution