Research map / Image generation / Editing and style transfer

Style Nursing with Spatial and Semantic Guidance for Zero-Shot Traffic Scene Style Transfer

AAAI 2025 · Image generation › Editing and style transfer · approach: Arbitrary neural style transfer · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a training-free zero-shot traffic-scene style transfer method on Stable Diffusion (Style Nursing plus Spatial and Semantic Guidance) and a 6K-image benchmark, STREET-6K.

Problem

Zero-shot text-guided style transfer with diffusion models neglects style words in prompts (Style Neglection) and struggles to preserve scene content in complex traffic scenes.

Approach

Maximizes cross-attention activation of style tokens during denoising (Style Nursing), and guides latents with spatial features from coarse U-Net layers plus self-attention maps of a SAM-segmented semantic image (Spatial and Semantic Guidance), all via gradient updates without training.

Key results

• Ours (full) achieves CLIP text-image score 92.6% and PSNR 25.28 on STREET-6K, with best or near-best content preservation; in a 70-participant user study Style Fidelity score 0.759 vs 0.149 for Plug-and-Play.
• Fine-tuning Yolop on STREET-6K extreme-weather images raises snowy/rainy/foggy Recall (e.g., foggy 52.4 to 74.7).
• Abstract claims Recall and mIoU improve by over 10% and 3% respectively in rarely-seen traffic scenes.

Measured effects reported in the paper

Surprising findings

• Ablation without Style Nursing scores higher on several content-preservation metrics (PSNR 26.56, SSIM 91.7%) than the full method, reflecting a trade-off with edit strength.

Limitations

• Quantitative evaluation of text-based editing lacks ground truth, relying on proxy metrics and a 70-person user study on 50 images.
• Spatial features from input image can leak appearance information, so deep layers are excluded, trading off style transfer and content preservation.

Experimental setup

Datasets: KITTI nuScenes ImageNet STREET-6K

Benchmarks: STREET-6K

Metrics: Structure Distance PSNR LPIPS MSE SSIM CLIP image similarity CLIP text-image score Recall mAP50 mIoU

Baselines: Plug-and-Play Prompt-to-Prompt InstructPix2Pix DiffuseIT Pix2Pix-zero SDEdit InstructDiffusion ControlNet T2I-Adapter DreamBooth

Builds on: Stable Diffusion Plug-and-Play Prompt-to-Prompt DDIM inversion Attend-and-Excite

Closest papers by meaning

• Z*: Zero-shot Style Transfer via Attention Reweighting (CVPR 2024 · similarity 0.86)
• Semantix: An Energy-guided Sampler for Semantic Style Transfer (ICLR 2025 · similarity 0.86)
• SigStyle: Signature Style Transfer via Personalized Text-to-Image Models (AAAI 2025 · similarity 0.86)
• SCSA: A Plug-and-Play Semantic Continuous-Sparse Attention for Arbitrary Semantic Style Transfer (CVPR 2025 · similarity 0.86)
• Style Injection in Diffusion: A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer (CVPR 2024 · similarity 0.86)
• CSGO: Content-Style Composition in Text-to-Image Generation (NeurIPS 2025 · similarity 0.86)