SEAL: Semantic-aware Single-image Sticker Personalization with a Large-scale Sticker-tag Dataset

2026-04-29 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

SEAL (SEmantic-aware single-image sticker personALization) is a new plug-and-play adaptation module designed to improve diffusion-based personalized text-to-image generation, specifically for sticker personalization. It addresses common issues like visual entanglement and structural rigidity that arise when using single reference images with test-time fine-tuning (TTF) methods. SEAL integrates into existing U-Net-based diffusion pipelines without modifying their backbones, employing a Semantic-guided Spatial Attention Loss, a Split-merge Token Strategy, and Structure-aware Layer Restriction during embedding adaptation. To facilitate this, the authors also introduce StickerBench, a large-scale sticker image dataset featuring structured tags across six attributes: Appearance, Emotion, Action, Camera Composition, Style, and Background. Experiments demonstrate that SEAL enhances identity preservation and contextual controllability by applying explicit spatial and structural constraints.

Key takeaway

For research scientists developing personalized text-to-image generation models, particularly for stickers, you should consider integrating SEAL into your existing diffusion pipelines. This module directly addresses common overfitting issues like visual entanglement and structural rigidity, improving both identity preservation and contextual controllability. Utilizing the accompanying StickerBench dataset can also provide a robust framework for evaluating attribute-level control and disentanglement in your models.

Key insights

SEAL improves single-image sticker personalization in diffusion models by mitigating overfitting and enhancing contextual control.

Principles

• Explicit spatial and structural constraints improve identity preservation.
• Structured datasets enable systematic evaluation of identity disentanglement.

Method

SEAL integrates into diffusion pipelines via a Semantic-guided Spatial Attention Loss, Split-merge Token Strategy, and Structure-aware Layer Restriction during embedding adaptation.

In practice

• Use StickerBench for sticker personalization research.
• Apply SEAL to reduce visual entanglement in single-image generation.

Topics

• Sticker Personalization
• Diffusion Models
• Single-Image Generation
• SEAL Module
• StickerBench Dataset

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Diff-SBSR: Learning Multimodal Feature-Enhanced Diffusion Models for Zero-Shot Sketch-Based 3D Shape Retrieval — Diffusion models, enhanced with multimodal features, can effectively perform zero-shot sketch-based 3D shape retrieval.
• Direct 3D-Aware Object Insertion via Decomposed Visual Proxies — DIRECT enables 3D-aware object insertion by decomposing guidance into appearance, geometry, and context, injected separately for precise control.
• LEXIS: LatEnt ProXimal Interaction Signatures for 3D HOI from an Image — LEXIS-Flow reconstructs 3D human-object interactions from images using dense proximity fields and learned interaction signatures.
• Rethinking Token Reduction for Diffusion Models via Output-Similarity-Awareness — DiTo improves Diffusion Transformer efficiency by focusing token reduction on output similarity, not just input similarity.
• ProductConsistency: Improving Product Identity Preservation in Instruction-Based Image Editing via SFT and RL — A new dataset and RL reward significantly improve product identity preservation and text fidelity in instruction-based image editing.
• Semantic-Structural Alignment for Generative Pictorial Charts — A dual-conditioned diffusion framework unifies semantic expression with structural fidelity for automated pictorial chart generation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.