TP-Blend: Textual-Prompt Attention Pairing for Precise Object-Style Blending in Diffusion Models

The document introduces TP-Blend, a novel, training-free framework designed to address the challenges of simultaneously introducing new objects and styles in text-conditioned diffusion models. Existing methods often struggle with precise object blending and fine-grained style transfer, especially in preserving high-frequency textural details. TP-Blend overcomes these limitations by utilizing two distinct textual prompts: one for the blend object and another for the target style, injecting both into a single denoising trajectory. The framework employs two core components: Cross-Attention Object Fusion (CAOF) and Self-Attention Style Fusion (SASF). CAOF leverages an optimal transport problem to integrate blend-object features for seamless morphological transitions, while SASF injects intricate, brush-stroke-level style via Detail-Sensitive Instance Normalization and context-aware Key/Value matrix substitution. This dual-prompt mechanism ensures precise content representation and faithful style transfer without interference, offering fine-grained control over both blending strength and texture. Extensive experiments demonstrate that TP-Blend generates high-resolution, photo-realistic edits with superior quantitative fidelity, perceptual quality, and inference speed compared to recent baselines. Its ability to unify object replacement, blending, and style transfer within one process enhances controllability without additional computational cost.
#TPBlend #DiffusionModels #ObjectStyleBlending #TextConditionedEditing #ImageEditing #AI #DeepLearning #GenerativeAI #StyleTransfer #ObjectFusion

paper - https://arxiv.org/pdf/2601.08011v1
subscribe - https://t.me/arxivpaper
donations:
USDT: 0xAA7B976c6A9A7ccC97A3B55B7fb353b6Cc8D1ef7
BTC: bc1q8972egrt38f5ye5klv3yye0996k2jjsz2zthpr
ETH: 0xAA7B976c6A9A7ccC97A3B55B7fb353b6Cc8D1ef7
SOL: DXnz1nd6oVm7evDJk25Z2wFSstEH8mcA1dzWDCVjUj9e
created with NotebookLM