🎨 About This Repository

A curated collection of style transfer research papers spanning 🖼️ 2D image, 🎬 video, 🧊 3D (NeRF / mesh / point cloud / 3DGS) and ⏳ 4D (dynamic scene) stylization — from classic neural style transfer to the latest diffusion- and foundation-model-era methods — organized by the Style-Carrier Taxonomy from our survey Style Transfer: A Decade Survey (revised version coming soon).

Domain tags: 🖼️ Image — single-image stylization · 🎬 Video — temporally consistent stylization · 🧊 3D — stylization of neural / explicit 3D representations · ⏳ 4D — stylization of dynamic, time-varying 3D scenes · 🌐 Multiple — methods covering more than one domain

Instead of grouping papers by year, backbone (CNN / GAN / Diffusion / DiT), or output domain, we ask a single question:

In what form is style-specific information materialized, and how is it reused for new content?

Every method has exactly one canonical home — five style carriers spanning 2015–2026 across 🖼️ image, 🎬 video, 🧊 3D and ⏳ 4D stylization.

🖼️ Style Transfer Examples

🧭 The Style-Carrier Taxonomy

📌 What the taxonomy is not based on: year, backbone family, output domain, or auxiliary mechanisms (ControlNet, SDS, DDIM inversion …) — those are tags, not categories. The same perceptual objective underlies Gatys (🎯 objective), Johnson (⚙️ parameter) and Ghiasi (🔌 condition); AdaIN, StyTr² and StyleAligned are all 🧩 feature operators despite spanning CNN, Transformer and Diffusion eras.

📊 Coverage at a Glance

Empty cells are open problems, not omissions — e.g. ⏳ 4D stylization remains nearly untouched.

📑 Table of Contents

1. 🎯 Style as Optimization Objective — 20 papers
2. 🧩 Style as Feature Operator — 27 papers
3. 🌀 Style as Latent Variable — 11 papers
4. ⚙️ Style as Model Parameters — 25 papers
5. 🔌 Style as Learned Condition — 27 papers
6. 📚 Datasets
7. 🤝 Contributing

🗂️ Papers by Style Carrier

1. 🎯 Style as Optimization Objective

Style exists only as a loss, score, or guidance signal — no transferable style representation is ever materialized.

2. 🧩 Style as Feature Operator

Style acts through direct matching, replacement, transport, or attention coupling of content/style activations.

3. 🌀 Style as Latent Variable

Style is a native generative latent with its own prior — sampleable, mixable, and editable without any reference.

4. ⚙️ Style as Model Parameters

Style is materialized as a persistent parameter asset (weights, LoRA, learned tokens) reusable across new content.

5. 🔌 Style as Learned Condition

An external reference is encoded into a standalone style condition consumed via a dedicated conditioning interface.

📚 Datasets

🖼️ Image Datasets

Dataset	Year	Size	Description	Link
WikiArt	2018	42,129 images	Style, Artist, Genre	link
Stylized ImageNet	2018	~134GB	Style Transfer	link
BAM	2017	2.5M images	Artistic Media	link
ArtBench-10	2022	60,000 images	Artwork Benchmark	link
FFHQ	2019	70,000 images	Human Faces	link
MetFaces	2020	1,336 images	Artistic Faces	link
AAHQ	2021	25,000 images	Artistic Faces	link
Ukiyo-e Faces	2020	5,209 images	Aligned Ukiyo-e Faces	link
DiffusionDB	2022	14M images	Text-to-image	link
JourneyDB	2023	4.4M images	Multimodal Vision	link
StyleShot	2024	-	Style Transfer	link
Danbooru2017	2017	2.94M images	Anime	link
Chinese Style Transfer	2018	1000 content, 100 style	Chinese Painting	link
4SKST	2023	25 color, 100 sketches	Sketch Style	link

🎬 Video Datasets

Dataset	Year	Size	Description	Link
UADFV	2018	100 videos	Video Style Transfer	link
FaceForensics++	2019	6000 videos	Swapped Face	link
Celeb-DF	2020	408 original videos	DeepFake	link
DFDC	2020	100,000 clips	DeepFake Detection	link
FFIW-10K	2021	10,000 videos	Face Forgery	link
ForgeryNet	2021	221,247 videos	Forgery Analysis	link

🧊 3D / Motion Datasets

Dataset	Year	Size	Description	Link
100STYLE	2022	4M frames	Stylized Motion Capture	link
Motiondataset	2023	36,673 frames	3D Motion	link

🤝 Contributing

We welcome contributions! When suggesting a paper, please indicate which style carrier it belongs to:

• 🎯 Objective — style lives only in the loss / score / guidance
• 🧩 Feature — direct activation matching, transport, or attention coupling
• 🌀 Latent — native generative latent, sampleable without a reference
• ⚙️ Parameter — persistent style asset (weights / LoRA / tokens)
• 🔌 Condition — reference-encoded standalone style condition

Ways to help:

• 🐛 Report bugs and issues
• 💡 Suggest new papers or resources
• 🔧 Submit pull requests
• ⭐ Star this repository if you find it helpful!

📖 Citation

If you find this repository useful for your research, please consider citing:

@article{zhang2025style,
  title={Style transfer: A decade survey},
  author={Zhang, Tianshan and Tang, Hao},
  journal={arXiv preprint arXiv:2506.19278},
  year={2025}
}

⭐ Star History

🙏 Acknowledgments

Thanks to all researchers and developers who made their work publicly available.