GraftSR

Grafting Authentic Textures for Real-World Image Super-Resolution via Identical-Instance Guidance

Qifan Yu¹, Haoran Bai¹, Zongyao He¹, Weijie He¹, Sibin Deng¹, Honggang Qi², Ying Chen^1,* ¹Taobao & Tmall Group of Alibaba ²University of Chinese Academy of Sciences {yuqifan.yqf, baihaoran.bhr}@taobao.com · ^*Corresponding author

📄 Paper ⌘ Code ◫ Dataset ↕ Results

Authentic Texture Restoration

Drag the slider to compare the low-resolution input with the high-resolution result restored by GraftSR.

Low Resolution

GraftSR

Example 01

Drag horizontally to compare · Use side buttons to switch examples

About the project

Abstract

Diffusion-based real-world image super-resolution (SR) achieves impressive perceptual quality but inherently suffers from severe texture hallucination. To overcome this limitation, we propose GraftSR, a texture-reference-guided generative SR framework that leverages reference images of the identical instance to anchor the restoration of authentic textures. However, severe spatial misalignment between low-quality inputs and their references poses significant challenges, often leading to ambiguous transfer targets and background feature leakage. To address these issues, GraftSR employs a novel dual-mask reference guidance mechanism that systematically decouples the cross-view texture injection process. By explicitly isolating what authentic textures to extract from the reference and precisely localizing where to apply them within the target, GraftSR achieves robust texture transfer without relying on brittle spatial alignment. Furthermore, to bridge the critical gap in appropriate training data, we construct TexRefSR-141K, the first large-scale dataset providing high-quality reference tuples equipped with complementary spatial masks. Extensive experiments on our newly established benchmark, TexRefSR-Eval, demonstrate that GraftSR sets a new state-of-the-art. Notably, it reduces LPIPS by 20.2% over top-performing baselines, achieving superior reference-faithful restoration.

Figure comparing GraftSR with existing super-resolution methods

Texture-Reference-Guided SR

We propose GraftSR, a novel texture-reference-guided generative SR framework that suppresses generative hallucinations and achieves highly faithful texture transfer from cross-view identical-instance references.

TexRefSR-141K & Benchmark

We construct TexRefSR-141K and TexRefSR-Eval, the first large-scale dataset and benchmark suite providing high-quality identical-instance tuples equipped with complementary spatial masks.

State-of-the-Art Results

Extensive experiments demonstrate that GraftSR establishes a new state-of-the-art in reference-guided SR, bridging the gap between perceptual quality and texture fidelity for faithful texture restoration.

Method

Method Overview

Dual-Mask Reference-Guided Architecture

Built upon an MMDiT backbone, GraftSR introduces a novel dual-mask reference guidance mechanism that decouples cross-view texture injection into two explicit questions — what to reference and where to transfer — eliminating the reliance on brittle spatial alignment between the low-quality input and the identical-instance reference.

Overall architecture of GraftSR with dual-mask reference guidance

Overall architecture of GraftSR. Mask-modulated reference tokens and region-aware semantic tokens jointly condition cascaded MMDiT blocks to denoise the dual-noise latent, which is distilled into an efficient one-step SR model.

Semantic & Spatial Condition Extraction

A vision-language model jointly analyzes I_LQ and I_REF to identify the shared instance and generate a unified caption T. SAM3 then performs zero-shot segmentation to obtain the target mask M_TGT and the reference mask M_REF, explicitly localizing the identical instance across both views.

Mask-Modulated Reference Tokens · What

I_REF is encoded by a frozen VAE, and its latent is additively modulated by the downsampled reference mask. This soft intervention isolates authentic texture-bearing regions from irrelevant backgrounds without disrupting the continuous latent space. z_ref = E_VAE(I_REF) + φ(D(M_REF))

Region-Aware Semantic Tokens · Where

A frozen Qwen2.5-VL encoder jointly embeds the low-quality input, the target mask, the reference, and the caption into region-aware semantic tokens that designate precisely where the extracted textures should be applied. z_sem = E_VLM(I_LQ, M_TGT, I_REF, T)

One-Step Adversarial Distillation

A dual-noise strategy balances perceptual quality and texture fidelity. The noisy target tokens are concatenated with the extracted conditions along the sequence dimension, and the model is trained as a one-step SR network optimized with MSE, DISTS, and a relativistic GAN loss driven by a Wan2.1-1.3B discriminator.

Identical-Instance Dataset Construction

Prevailing reference-based SR datasets merely pair semantically approximate images. We instead exploit large-scale e-commerce multi-view galleries, which uniquely capture strictly identical instances across diverse viewpoints, and build TexRefSR-141K through an automated VLM-driven curation pipeline.

VLM-driven data curation pipeline for TexRefSR-141K

The proposed VLM-driven data curation pipeline. It retrieves identical-instance triplets from multi-view galleries and extracts the complementary masks and unified caption to support authentic texture transfer.

Identical-Instance Triplet

Anchored on a quality-filtered HQ reference I_REF, the pipeline retrieves a multi-view HQ ground truth I_HQ of the same instance and synthesizes I_LQ via a high-order degradation model — forming the core {I_REF, I_HQ, I_LQ} triplet with genuine texture correspondence rather than mere semantic resemblance.

Caption & Complementary Masks

The same condition extraction annotates each triplet with a unified caption T and the complementary mask pair {M_REF, M_TGT}. Both are derived from the clearer I_HQ, and random morphological dilation keeps delicate boundaries (e.g., lace and decorative edges) robust.

Rigorous Quality Assessment

Candidates are filtered before masking: embedding similarity discards mismatched pairs, a quality predictor eliminates degraded samples, and extreme close-ups that lose recognizable instance identity are rejected — yielding over 141K high-quality tuples across 61K diverse fashion products.

1-to-N Mapping & TexRefSR-Eval

Each reference maps to multiple target ground truths, yielding rich cross-view texture correspondences. We further curate TexRefSR-Eval with 300 manually verified cases: 250 synthesized samples for full-reference evaluation and 50 naturally degraded real-world samples.

Evaluation

Quantitative Results

Type	Method	Full-Reference				No-Reference
Type	Method	LPIPS ↓	DISTS ↓	PSNR ↑	SSIM ↑	MUSIQ ↑	NIQE ↓	CLIPIQA ↑	MANIQA ↑
Open-source Generative SR	OSEDiff	0.2326	0.1444	28.3752	0.8471	65.2201	4.8256	0.6631	0.5250
	SUPIR	0.2208	0.1321	29.3151	0.8382	61.5577	5.2317	0.5922	0.5225
	DiT4SR	0.3816	0.2147	25.6121	0.8071	33.1845	7.3649	0.5769	0.4253
	ODTSR	0.1983	0.1048	29.5098	0.8431	69.2601	4.7789	0.6176	0.5129
	PiSA-SR	0.2247	0.1343	28.5478	0.8458	66.7754	4.6765	0.6831	0.5338
	VOSR-1.4b	0.2349	0.1247	28.3661	0.8290	71.1585	4.3381	0.6873	0.5523
Open-source Reference-based SR	RefSR	0.4384	0.2685	28.6190	0.8297	24.0680	8.1108	0.4106	0.3259
	AdaRefSR	0.6071	0.2785	16.2155	0.7154	32.1215	7.7052	0.5645	0.4367
	GarmentZoom	0.4156	0.2643	25.8811	0.7255	34.1616	6.1379	0.4360	0.4160
Commercial Editing Models	GPT-Image-2	0.4474	0.2154	18.3255	0.7665	34.4500	7.4755	0.5781	0.4355
Commercial Editing Models	Gemini-3-Pro	0.2065	0.1034	26.7874	0.8222	64.9845	4.9920	0.6216	0.5048
	Graft (Ours)	0.1583	0.0986	30.1060	0.8494	69.8535	4.9346	0.7275	0.5270

Quantitative comparison on the TexRefSR-Eval-Syn benchmark, where ground-truth references are available. Both full-reference and no-reference metrics are reported. Red bold and blue underline denote the best and the second-best results.

Type	Method	No-Reference				VLM Evaluation (GPT-4.1-global)
Type	Method	NIQE ↓	MUSIQ ↑	CLIPIQA ↑	MANIQA ↑	PQ ↑	TC ↑	Overall ↑
Open-source Generative SR	OSEDiff	3.9329	70.5668	0.6395	0.6292	3.320	1.740	2.530
	SUPIR	4.2105	59.6691	0.4778	0.5741	3.380	1.960	2.670
	DiT4SR	3.7190	72.5782	0.6428	0.6485	3.720	2.040	2.880
	ODTSR	4.2809	71.5261	0.6140	0.6580	2.760	1.560	2.160
	PiSA-SR	4.2523	65.5855	0.4913	0.5789	2.840	1.660	2.250
	VOSR-1.4b	3.8273	69.7636	0.6227	0.6399	3.880	2.380	3.130
Open-source Reference-based SR	RefSR	5.0782	53.8617	0.3251	0.5030	2.200	1.280	1.740
	AdaRefSR	5.7189	46.2674	0.3625	0.5087	2.580	1.560	2.070
	GarmentZoom	5.6298	31.4101	0.3625	0.4417	2.380	1.840	2.110
Commercial Editing Models	GPT-Image-2	4.9825	65.9262	0.5667	0.6159	3.020	2.100	2.560
Commercial Editing Models	Gemini-3-Pro	5.1077	56.5937	0.4633	0.5277	3.880	2.700	3.290
	GraftSR (Ours)	3.9881	73.0045	0.6409	0.6507	3.980	2.840	3.410

Quantitative comparison on the TexRefSR-Eval-Real benchmark with no-reference metrics and VLM-based metrics.

Method	PSNR ↑	SSIM ↑	LPIPS ↓	MUSIQ ↑	MANIQA ↑
SUPIR	23.65	0.6620	0.3541	62.09	0.5780
DiT4SR	23.50	0.6657	0.3215	67.76	0.6564
OSEDiff	25.15	0.7341	0.2920	69.08	0.6335
PiSA-SR	25.40	0.7418	0.2672	70.14	0.6551
VOSR-1.4B	25.23	0.7175	0.2732	70.58	0.6443
ODTSR	25.07	0.7361	0.2398	68.29	0.6622
GraftSR (Ours)	25.51	0.7512	0.2559	70.37	0.6676

RealSR

Method	PSNR ↑	SSIM ↑	LPIPS ↓	MUSIQ ↑	MANIQA ↑
SUPIR	25.09	0.6460	0.4243	58.79	0.5471
DiT4SR	25.40	0.6657	0.3869	65.75	0.6287
OSEDiff	27.92	0.7835	0.2968	64.65	0.5895
PiSA-SR	28.32	0.7804	0.2960	66.11	0.6146
VOSR-1.4B	27.88	0.7413	0.3260	66.04	0.6053
ODTSR	28.14	0.7736	0.2592	62.86	0.6227
GraftSR (Ours)	29.04	0.8052	0.2858	66.73	0.6376

DRealSR

Quantitative comparison on general real-world SR. Our GraftSR still achieves strong generalization in both fidelity and perceptual quality even without reference images.

Visualizations

More Qualitative Results

More qualitative comparisons on the TexRefSR-Eval-Syn dataset, where low-quality inputs are produced by the synthetic degradation model. Guided by the reference, our method restores textures that are more faithful to the ground truth, recovering fine patterns and structural details that other methods either over-smooth or hallucinate incorrectly.

More qualitative comparisons on the TexRefSR-Eval-Real dataset, which contains real-world low-quality images without ground truth. Our method produces perceptually sharper results whose textures remain consistent with the reference, whereas other methods tend to yield blurry or unfaithful reconstructions.

More qualitative comparisons on general real-world SR benchmarks (RealSR and DRealSR). Although no reference is available in this setting, our method still generalizes well to natural scenes and recovers clean, realistic details that are competitive with or superior to state-of-the-art general SR methods.