Research map / Image generation / Fast diffusion sampling
OCD: Learning to Overfit with Conditional Diffusion Models
Shahar Lutati, Lior Wolf
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 |
|---|---|---|---|
| Apply OCD to LeNet5 | LeNet5 · MNIST test | accuracy: 99.2 ± 0.1 -> 99.7 ± 0.1 | Improved |
| Apply OCD to GoogleNet | GoogleNet · CIFAR10 test | accuracy: 92.85 ± 0.40 -> 93.68 ± 0.38 | Improved |
| Use an ensemble of five OCD samples (ablation) | DeiT-B/16-D + OCD · TinyImageNet test | accuracy: 90.80 ± 0.35 -> 92.00 ±0.6 | Improved |
| Use a fixed global scale instead of input-dependent scaling (ablation) | OCD · MNIST test | accuracy: 99.7 ± 0.1 -> 99.3 ± 0.1 | Hurt |
| Use a fixed global scale instead of input-dependent scaling (ablation) | OCD · Lego | MSE: .008 ± .006 -> .009 ± .008 | Hurt |
| Use a fixed global scale instead of input-dependent scaling (ablation) | OCD · California Housing | MSE: 0.4800 ± 0.0020 -> 0.5000 ± 0.0030 | Hurt |
| Use a fixed global scale instead of input-dependent scaling (ablation) | OCD · Libri5Mix | SI-SDRi[dB]: 12.8 ± 0.3 -> 13.4 ± 0.1 | Hurt |
| Replace OCD generation with the nearest training sample's overfit model (ablation) | OCD · CIFAR10 test | accuracy: 93.68 ± 0.38 -> 93.03 ± 0.40 | Hurt |
| Choose the alternative layer instead of the highest-scoring layer (ablation) | OCD · CIFAR10 test | accuracy: 93.68 ± 0.38 -> 92.96± 0.27 | Hurt |
| Apply OCD to the SetFit model | SetFit · Mean across few-shot NLP datasets | accuracy: 62.2 -> 64.3 | Improved |
| Use an ensemble of five OCD inferences (ablation) | SetFit + OCD · Mean across few-shot NLP datasets | accuracy: 64.3 -> 64.8 | Improved |
| Average the weights of five OCD networks instead of using one network (ablation) | SetFit + OCD · Few-shot NLP classification | accuracy | No clear effect |
Closest papers by meaning
- Random Conditioning for Diffusion Model Compression with Distillation (CVPR 2025 · similarity 0.89)
- APT: Adaptive Personalized Training for Diffusion Models with Limited Data (CVPR 2025 · similarity 0.88)
- CleanDIFT: Diffusion Features without Noise (CVPR 2025 · similarity 0.88)
- CADS: Unleashing the Diversity of Diffusion Models through Condition-Annealed Sampling (ICLR 2024 · similarity 0.88)
- Label-Noise Robust Diffusion Models (ICLR 2024 · similarity 0.87)
- Your Diffusion Model is Secretly a Noise Classifier and Benefits from Contrastive Training (NeurIPS 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.
