Project Page
SEAL logo
SEAL : Semantic-aware Single-image Sticker
Personalization with a Large-scale Sticker-tag Dataset
Changhyun Roh, Yonghyun Jeong, Jonghyun Lee, Chanho Eom*, Jihyong Oh*
Chung-Ang University NAVER Cloud Lunit Inc.
*Co-corresponding Authors
Paper PDF Code Dataset BibTeX
TL;DR: A plug-and-play semantic adaptation module for single-image sticker personalization, designed to mitigate visual entanglement and structural rigidity in test-time fine-tuning pipelines.
CoRe-based Results
Reference
Input 1
Output 1
Output 1-1
Output 2
Output 1-2
S*, sitting at desk, hands on table, sticker style, green background
Reference
Input 2
Output 1
Output 2-1
Output 2
Output 2-2
S*, standing, close-up front view, animation style, cityscape background
Reference
Input 3
Output 1
Output 3-1
Output 2
Output 3-2
S*, front view, anime style, forest background
UnZipLoRA-based Results
Reference
Input 4
Output 1
Output 4-1
Output 2
Output 4-2
S*, leaning over bed, sticker style, room background
Reference
Input 5
Output 1
Output 5-1
Output 2
Output 5-2
S*, arms stretched, front view, sticker style, room background
Reference
Input 6
Output 1
Output 6-1
Output 2
Output 6-2
S*, holding canned food, close-up front view, sticker style, cityscape background
SEAL generated results overview

Synthesizing a target concept from a single reference image remains challenging in diffusion-based personalized text-to-image generation, particularly when prompts require explicit attribute edits. In the sticker domain, test-time fine-tuning methods often overfit to the reference image and suffer from visual entanglement and structural rigidity. We introduce SEAL, a plug-and-play, architecture-agnostic adaptation module that combines a Semantic-guided Spatial Attention Loss, a Split-merge Token Strategy, and a Structure-aware Layer Restriction. We also introduce StickerBench, a large-scale sticker dataset with six structured attributes for controlled evaluation of identity disentanglement and contextual controllability in single-image sticker personalization.

Method

SEAL is a plug-and-play module for existing personalization pipelines.

SEAL method overview
1

Semantic-guided Spatial Attention Loss

Aligns the concept-token cross-attention map with the object region predicted by SAM, suppressing background leakage and improving identity disentanglement.

2

Split-merge Token Strategy

Optimizes multiple auxiliary embeddings and merges them into one concept embedding, improving optimization stability under single-image supervision.

3

Structure-aware Layer Restriction

Applies spatial supervision only to semantically informative cross-attention layers, reducing overfitting to low-level layout patterns.

Seven layer-position metrics across SD2.1 and SDXL
Layer selection across backbones. Intermediate injection positions balance identity preservation, attention alignment, background leakage, and prompt alignment in SD2.1 and SDXL. Open the figure for full resolution.
Key Insights

Structural Rigidity and Background Entanglement

SEAL is designed to reduce the two dominant failure modes in single-image sticker personalization.

Structural Rigidity

Existing methods often memorize the reference-specific layout, which reduces flexibility under action, pose, and composition edits.

Input Reference
REFERENCE
:
Baseline result with problem
BASELINE
+ SEAL
SEAL result
Ours

Background Entanglement

Existing methods may absorb background cues into the learned concept representation, causing identity and context to become entangled.

Input Reference
REFERENCE
:
Baseline result with problem
BASELINE
+ SEAL
SEAL result
Ours
Evidence

Results

Quantitative comparisons from the revised manuscript, paired with visual analyses of SEAL's adaptation behavior.

StickerBench · 200 concepts · nine methods

Main comparison

Each cell compares a baseline with its SEAL-integrated variant. Scores are averaged across 20 prompts and 10 inference seeds per concept.

Prompt alignment
StickerBench prompt alignment
MethodCLIP-T ↑CycleReward ↑
Textual Inversion
0.2710.296(+0.025)
95% CI [0.259, 0.283] → [0.284, 0.308]
0.4760.487(+0.011)
95% CI [0.465, 0.487] → [0.476, 0.498]
DreamBooth
0.2910.311(+0.020)
95% CI [0.279, 0.303] → [0.300, 0.322]
0.4730.498(+0.025)
95% CI [0.460, 0.486] → [0.485, 0.511]
AttnDreamBooth
0.2570.286(+0.029)
95% CI [0.247, 0.267] → [0.277, 0.295]
0.4880.476(-0.012)
95% CI [0.477, 0.499] → [0.465, 0.487]
ClassDiffusion
0.2910.288(-0.003)
95% CI [0.288, 0.294] → [0.285, 0.291]
0.4820.499(+0.017)
95% CI [0.471, 0.493] → [0.486, 0.512]
Custom Diffusion
0.2980.316(+0.018)
95% CI [0.288, 0.308] → [0.307, 0.325]
0.4960.507(+0.011)
95% CI [0.485, 0.507] → [0.496, 0.518]
CoRe
0.2760.289(+0.013)
95% CI [0.267, 0.285] → [0.280, 0.298]
0.5340.551(+0.017)
95% CI [0.523, 0.545] → [0.539, 0.563]
UnZipLoRA
0.3380.344(+0.006)
95% CI [0.331, 0.345] → [0.336, 0.352]
0.6210.650(+0.029)
95% CI [0.610, 0.632] → [0.640, 0.660]
T-LoRA
0.3120.338(+0.026)
95% CI [0.302, 0.322] → [0.329, 0.347]
0.5920.622(+0.030)
95% CI [0.580, 0.604] → [0.611, 0.633]
TextBoost
0.2910.298(+0.007)
95% CI [0.284, 0.298] → [0.290, 0.306]
0.5200.538(+0.018)
95% CI [0.508, 0.532] → [0.526, 0.550]
Identity preservation
StickerBench identity preservation
MethodCLIP-I ↑DINOv2 ↑CP ↑VSM ↑
Textual Inversion
0.6850.671(-0.014)
95% CI [0.670, 0.700] → [0.657, 0.685]
0.4030.437(+0.034)
95% CI [0.369, 0.437] → [0.403, 0.471]
0.5650.569(+0.004)
95% CI [0.561, 0.569] → [0.565, 0.573]
0.7980.800(+0.002)
95% CI [0.796, 0.800] → [0.798, 0.802]
DreamBooth
0.5900.628(+0.038)
95% CI [0.566, 0.614] → [0.604, 0.652]
0.3040.421(+0.117)
95% CI [0.267, 0.341] → [0.383, 0.459]
0.5790.594(+0.015)
95% CI [0.564, 0.594] → [0.580, 0.608]
0.8060.798(-0.008)
95% CI [0.798, 0.814] → [0.790, 0.806]
AttnDreamBooth
0.7710.783(+0.012)
95% CI [0.758, 0.784] → [0.771, 0.795]
0.5330.597(+0.064)
95% CI [0.500, 0.566] → [0.569, 0.625]
0.5680.587(+0.019)
95% CI [0.555, 0.581] → [0.574, 0.600]
0.8200.822(+0.002)
95% CI [0.818, 0.822] → [0.820, 0.824]
ClassDiffusion
0.7510.790(+0.039)
95% CI [0.729, 0.773] → [0.769, 0.811]
0.6560.691(+0.035)
95% CI [0.621, 0.691] → [0.656, 0.726]
0.6020.638(+0.036)
95% CI [0.586, 0.618] → [0.622, 0.654]
0.8280.852(+0.024)
95% CI [0.819, 0.837] → [0.842, 0.862]
Custom Diffusion
0.8310.859(+0.028)
95% CI [0.814, 0.848] → [0.840, 0.878]
0.6700.707(+0.037)
95% CI [0.641, 0.699] → [0.678, 0.736]
0.5980.614(+0.016)
95% CI [0.586, 0.610] → [0.601, 0.627]
0.8350.838(+0.003)
95% CI [0.831, 0.839] → [0.835, 0.841]
CoRe
0.7910.838(+0.047)
95% CI [0.774, 0.808] → [0.821, 0.855]
0.5120.579(+0.067)
95% CI [0.482, 0.542] → [0.551, 0.607]
0.6280.665(+0.037)
95% CI [0.616, 0.640] → [0.652, 0.678]
0.8170.854(+0.037)
95% CI [0.816, 0.818] → [0.831, 0.877]
UnZipLoRA
0.8360.859(+0.023)
95% CI [0.818, 0.854] → [0.841, 0.877]
0.6940.712(+0.018)
95% CI [0.672, 0.716] → [0.691, 0.733]
0.7140.723(+0.009)
95% CI [0.703, 0.725] → [0.711, 0.735]
0.9210.939(+0.018)
95% CI [0.912, 0.930] → [0.931, 0.947]
T-LoRA
0.7120.725(+0.013)
95% CI [0.698, 0.726] → [0.712, 0.738]
0.4060.461(+0.055)
95% CI [0.372, 0.440] → [0.426, 0.496]
0.5880.616(+0.028)
95% CI [0.573, 0.603] → [0.602, 0.630]
0.9340.952(+0.018)
95% CI [0.925, 0.943] → [0.943, 0.961]
TextBoost
0.7760.790(+0.014)
95% CI [0.761, 0.791] → [0.774, 0.806]
0.5280.531(+0.003)
95% CI [0.524, 0.532] → [0.527, 0.535]
0.6360.642(+0.006)
95% CI [0.630, 0.642] → [0.636, 0.648]
0.8920.910(+0.018)
95% CI [0.889, 0.895] → [0.894, 0.926]
Structural variation
StickerBench structural variation
MethodDINO-nSSM ↑
Textual Inversion
0.6130.633(+0.020)
95% CI [0.594, 0.632] → [0.615, 0.651]
DreamBooth
0.6270.661(+0.034)
95% CI [0.604, 0.650] → [0.639, 0.683]
AttnDreamBooth
0.6300.644(+0.014)
95% CI [0.616, 0.644] → [0.630, 0.658]
ClassDiffusion
0.6240.641(+0.017)
95% CI [0.605, 0.643] → [0.624, 0.658]
Custom Diffusion
0.6390.652(+0.013)
95% CI [0.624, 0.654] → [0.637, 0.667]
CoRe
0.6430.696(+0.053)
95% CI [0.624, 0.662] → [0.679, 0.713]
UnZipLoRA
0.6710.675(+0.004)
95% CI [0.666, 0.676] → [0.671, 0.679]
T-LoRA
0.6590.648(-0.011)
95% CI [0.645, 0.673] → [0.635, 0.661]
TextBoost
0.6510.673(+0.022)
95% CI [0.633, 0.669] → [0.652, 0.694]

Mean scores and 95% concept-level bootstrap confidence intervals are from the revised main comparison. DINO-nSSM is interpreted together with identity preservation and prompt alignment.

Base and SEAL results for five single-attribute edits across seven metrics and Qwen evaluation
Attribute-level prompt edits. Open circles are baselines and filled circles are their SEAL variants. E, A, C, S, and B stand for Emotion, Action, Camera Composition, Style, and Background.
Controlled attribute edits

Attention and generation

CoRe and CoRe + SEAL use matched reference images, prompts, and inference settings.

Reference images, generated images, and attention maps before and after background, action, and composition edits
Before and after one attribute changes. Generated images and subject and tag attention maps are shown for Background, Action, and Camera Composition edits.
Attention and generation metrics under single-attribute prompt edits
Edited attribute · MethodAttention leakage ↓Background similarityCLIP-T ↑CycleReward ↑
Background · CoRe0.5120.1350.2850.549
Background · +SEAL0.4890.0260.2950.559
Action · CoRe0.5030.1190.2690.520
Action · +SEAL0.4870.0620.2860.542
Camera Composition · CoRe0.4820.3530.2640.518
Camera Composition · +SEAL0.4530.1780.2800.538
Emotion · CoRe0.4240.0920.2810.546
Emotion · +SEAL0.4200.0530.2930.561
Style · CoRe0.5880.2780.2810.537
Style · +SEAL0.5500.2140.2910.555

Background similarity is lower-is-better only for the Background-edit condition. For other edits it has no preferred direction.

CustomConcept101 · 101 concepts

Cross-dataset evaluation

Five baselines and their SEAL variants are evaluated with the same reference, prompts, and inference settings within each pair.

Prompt alignment and structural variation
CustomConcept101 prompt alignment and structural variation
MethodCLIP-T ↑CycleReward ↑DINO-nSSM ↑
Custom Diffusion
0.2810.285(+0.004)
0.5090.530(+0.021)
0.6430.673(+0.030)
CoRe
0.2740.291(+0.017)
0.5460.535(-0.011)
0.6300.660(+0.030)
UnZipLoRA
0.3170.324(+0.007)
0.6170.604(-0.013)
0.6780.706(+0.028)
T-LoRA
0.3090.326(+0.017)
0.5980.617(+0.019)
0.6590.662(+0.003)
TextBoost
0.3040.297(-0.007)
0.5150.536(+0.021)
0.6640.691(+0.027)
Identity preservation
CustomConcept101 identity preservation
MethodCLIP-I ↑DINOv2 ↑CP ↑VSM ↑
Custom Diffusion
0.7480.797(+0.049)
0.5350.557(+0.022)
0.5970.599(+0.002)
0.8410.852(+0.011)
CoRe
0.7540.773(+0.019)
0.5080.562(+0.054)
0.6190.626(+0.007)
0.8300.849(+0.019)
UnZipLoRA
0.8030.864(+0.061)
0.6390.708(+0.069)
0.7040.725(+0.021)
0.9060.897(-0.009)
T-LoRA
0.7870.802(+0.015)
0.4510.476(+0.025)
0.5950.626(+0.031)
0.9290.937(+0.008)
TextBoost
0.7910.815(+0.024)
0.5390.564(+0.025)
0.6310.663(+0.032)
0.9000.909(+0.009)

Values are from the revised cross-dataset comparison. Changes across individual metrics can trade off; DINO-nSSM is read alongside identity and prompt scores.

Human and vision-language-model preference rates with correlations across three evaluation criteria
Human and VLM preference. The same 30 pairwise questions cover disentanglement, controllability, and overall quality. Scatterplots compare the VLM and human preference rates.
View the preference rates as bars
Human, GPT, and Qwen SEAL preference rates as grouped bar charts
The same human and VLM preference results in a grouped bar chart.
Component analysis · CoRe baseline

What each component contributes

All eight combinations of Spatial Attention Loss (SL), Split-merge Token Strategy (ST), and Layer Restriction (LR).

SEAL component and pairwise ablation
ComponentsCLIP-T ↑CycleReward ↑CLIP-I ↑DINOv2 ↑CP ↑DINO-nSSM ↑
None0.2760.5340.7910.5120.6280.643
SL0.2630.4560.8840.4810.5300.647
ST0.2960.4390.8210.4610.6000.651
LR0.2690.4840.9040.5750.5780.673
SL + ST0.2640.4970.8790.4190.6390.654
SL + LR0.2830.5360.8080.4800.6120.685
ST + LR0.2800.5230.7910.4630.6500.688
Full SEAL0.2890.5510.8380.5790.6650.696

The full configuration has the highest CycleReward, DINOv2, CP, and DINO-nSSM among these eight settings.

Split-merge strategy

Token count

The selected K = 5 balances the three reported metrics.

Split token count ablation
SettingCLIP-T ↑CLIP-I ↑DINOv2 ↑
K = 10.3130.8190.480
K = 30.3080.8330.483
K = 5 (selected)0.3060.8490.591
K = 70.2900.8840.479
Prompt representation

Tags at train and inference

Structured tags are most effective when used consistently.

Prompt representation ablation
Adaptation → inferenceCLIP-T ↑CLIP-I ↑DINOv2 ↑
Sentence → Sentence0.2610.8540.423
Sentence → Tag0.2580.8130.389
Tag → Sentence0.2730.5290.369
Tag → Tag0.3060.8490.591
More attribute analyses
Base and SEAL scores across human, animal, object, and fantasy-creature sticker categories
Appearance groups. Results are grouped by reference subject category.
Mean absolute metric changes for single-attribute prompt edits across five personalization methods
Magnitude of attribute changes. Mean absolute score changes are paired by concept and inference seed; shaded regions show one standard deviation.
Diffusion-loss trajectories for split-token merges at steps 50, 100, 150, and 250
Merge timing diagnostic. Diffusion-loss trajectories compare merge steps 50, 100, 150, and 250 under a 250-update budget.

Visual analyses

Qualitative ablations and attention maps from the manuscript.

Generated samples and attention maps for the SEAL component ablation
Component ablation on StickerBench using CoRe.
Structural rigidity at different spatial-supervision layers
Layer restriction and structural rigidity.
Cross-attention maps during adaptation and inference
Cross-attention maps during adaptation and inference.
Inference attention maps across split token counts
Inference attention behavior across split token counts.
Optimization stability for different split token counts
Optimization stability across split token counts.
Dataset

StickerBench

A curated sticker dataset with structured tags for controlled, single-image personalization evaluation.

261Kcurated sticker images

One identity, six structured attributes

Appearance anchors the concept; the remaining fields support controlled prompt edits.

01 · IDENTITYAppearanceTarget concept
02 · EXPRESSIONEmotionExpression and mood
03 · POSEActionWhat the subject does
04 · FRAMINGCamera CompositionView and crop
05 · RENDERINGStyleVisual treatment
06 · CONTEXTBackgroundSurrounding scene
Four sticker examples with six attribute tags, including two correct annotations and two action-tag errors
Annotation audit. Two examples have matching tags; two show incorrect action details marked in red. Character and View correspond to Appearance and Camera Composition in the six-attribute schema.

BibTeX

@misc{roh2026sealsemanticawaresingleimagesticker,
      title={SEAL: Semantic-aware Single-image Sticker Personalization with a Large-scale Sticker-tag Dataset}, 
      author={Changhyun Roh and Yonghyun Jeong and Jonghyun Lee and Chanho Eom and Jihyong Oh},
      year={2026},
      eprint={2604.26883},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.26883}, 
}