Research map / Probabilistic and scientific modeling / Variational and generative sampling
SCFlow: Implicitly Learning Style and Content Disentanglement with Flow Models
Pingchuan Ma, Xiaopei Yang, Yusong Li, Ming Gui, Felix Krause, Johannes Schusterbauer, Björn Ommer
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SCFlow embeddings instead of CLIP | CLIP embeddings · Style clustering, 51 styles | NMI: 0.4016 -> 0.8696 | Improved |
| use SCFlow embeddings instead of CLIP | CLIP embeddings · Content clustering, 200 contents | NMI: 0.5368 -> 0.8356 | Improved |
| use SCFlow embeddings instead of CSD | CSD style embeddings · Content clustering, 25 contents | NMI: 0.1888 -> 0.8374 | Improved |
| use SCFlow embeddings instead of CSD | CSD style embeddings · Style clustering, 51 styles | FDR: 0.6328 -> 3.5184 | Improved |
| use SCFlow embeddings instead of CLIP | CLIP embeddings · Style clustering, 51 styles | FDR: 0.2961 -> 3.5184 | Improved |
| use SCFlow content embeddings instead of CLIP | CLIP embeddings · ImageNet-1k kNN (k=1) | accuracy: 67.10 -> 66.25 | Hurt |
| use SCFlow style embeddings instead of CLIP | CLIP embeddings · WikiArt retrieval | Recall@1: 59.40 -> 65.34 | Improved |
| use SCFlow style embeddings instead of CSD-S | CSD-S · WikiArt retrieval | Recall@100: 95.58 -> 97.67 | Improved |
| train flow-matching merge model instead of contrastive loss on the same dataset (ablation) | contrastive loss baseline · Style clustering, 51 styles | NMI: 0.2905 -> 0.8696 | Improved |
| train flow-matching merge model instead of InfoNCE on the same dataset (ablation) | InfoNCE baseline · Style clustering, 51 styles | NMI: 0.5904 -> 0.8696 | Improved |
| train flow-matching merge model instead of InfoNCE on the same dataset (ablation) | InfoNCE baseline · WikiArt style clustering | NMI: 0.4238 -> 0.4010 | Hurt |
| train flow-matching merge model instead of InfoNCE on the same dataset (ablation) | InfoNCE baseline · ImageNet content clustering | FDR: 0.2056 -> 1.4264 | Improved |
Closest papers by meaning
- Simple Disentanglement of Style and Content in Visual Representations (ICML 2023 · similarity 0.91)
- Unsupervised Disentanglement of Content and Style via Variance-Invariance Constraints (ICLR 2025 · similarity 0.88)
- Disentangled Representation Learning via Flow Matching (NeurIPS 2026 · similarity 0.87)
- Differential Vector Erasure: Unified Training-Free Concept Erasure for Flow Matching Models (NeurIPS 2026 · similarity 0.87)
- Disentanglement Beyond Static vs. Dynamic: A Benchmark and Evaluation Framework for Multi-Factor Sequential Representations (NeurIPS 2025 · similarity 0.86)
- Factorized Diffusion Autoencoder for Unsupervised Disentangled Representation Learning (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
