Research map / Image generation / Text-to-image diffusion
DriftScope: Measuring The Hidden Effects of Diffusion Model Fine-Tuning
Héctor Laria, Yiping Han, Julian Santamaria, Kai WANG, Bogdan Raducanu, Joost Van de Weijer, Alex Gomez-Villa
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 |
|---|---|---|---|
| run concept unlearning (SPM) on nudity | SD1.4 · Zero-shot GDC CIFAR-10 | worst-case class accuracy change: -18.9 worst-case, FID 3.643 | Hurt |
| run concept unlearning (EraseDiff) | SD1.4 · Zero-shot GDC | accuracy / FID: FID 321.621; CIFAR-10 -91.9 | Hurt |
| fine-tune with DreamBooth (customization) | SD1.5 · Flowers-102 zero-shot | mean accuracy change: -2.88 ± 12.21 | Mixed |
| use DriftScope high-drift tokens instead of low-drift tokens (ablation) | SPM · DriftScope token evaluation | CLIP-i: 0.96 -> 0.88 | Improved |
| use DriftScope high-drift tokens instead of low-drift tokens (ablation) | SPM · DriftScope token evaluation | LPIPS: 0.14 -> 0.33 | Improved |
| use high-drift tokens instead of low-drift tokens (ablation) | ESD · DriftScope token evaluation | CLIP-i: 0.79 -> 0.72 | Mixed |
| use high-drift tokens instead of low-drift tokens (ablation) | DreamBooth SD3.5 · DriftScope token evaluation | Q-Eval: 0.71 -> 0.69 | No clear effect |
| use high-drift tokens instead of low-drift tokens (ablation) | Scissorhands · DriftScope token evaluation | Q-Eval: 0.19 -> 0.12 | Mixed |
| use L1 instead of L2 for attention divergence (ablation) | DriftScope · token localization | localization sharpness | Improved |
Closest papers by meaning
- The Hidden Language of Diffusion Models (ICLR 2024 · similarity 0.90)
- Concept Reachability in Diffusion Models: Beyond Dataset Constraints (ICML 2025 · similarity 0.89)
- You Can’t Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning (NeurIPS 2026 · similarity 0.89)
- CASL: Concept-Aligned Sparse Latents for Interpreting Diffusion Models (NeurIPS 2026 · similarity 0.88)
- Unveiling Concept Attribution in Diffusion Models (NeurIPS 2025 · similarity 0.88)
- CONFORM: Contrast is All You Need for High-Fidelity Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.87)
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.
