Research map / Image generation / Fast diffusion sampling
Transfer Learning for Diffusion Models
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 Transfer Guided Diffusion Process (TGDP), which adapts a source-domain diffusion model to a small target domain by adding a score guidance term estimated from a domain classifier's density ratio, rather than finetuning it.
Problem
Diffusion models need many training samples; with limited target-domain data, finetuning or regularizing a source-pretrained diffusion model overfits and degrades quality.
Approach
Shows the optimal target-domain score equals the frozen source score plus a guidance term log E_{p(x0|xt)}[q(x0)/p(x0)], which is estimated via a domain classifier density ratio and a guidance network trained with cycle and consistency regularizers; the source model is never finetuned.
Key results
- On Gaussian mixture with n=10 target samples, TGDP average likelihood 0.417 vs 0.290 for Finetune and 0.145 for Vanilla Diffusion.
- On ECG benchmark, TGDP achieves FID 10.46 and diversity 0.53 versus 11.01 and 0.37 for Vanilla Diffusion, with 2.8M trainable parameters vs 50.2M.
- Downstream ECG classification AUC of 0.953 for TGDP vs 0.941 for Finetune Generator and 0.906 for Vanilla Classifier.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TGDP instead of Finetune | Diffusion model transfer · Gaussian mixture, n=10 target samples | average likelihood: 0.290 -> 0.417 | Helped |
| Use TGDP instead of Finetune | Diffusion model transfer · Gaussian mixture, n=100 target samples | average likelihood: 0.329 -> 0.627 | Helped |
| Use TGDP instead of Finetune | Diffusion model transfer · Gaussian mixture, n=1000 target samples | average likelihood: 0.335 -> 0.673 | Helped |
| Use TGDP instead of Finetune Generator | ECG generation · ECG benchmark | Diversity: 0.47 -> 0.53 | Helped |
| Use TGDP instead of Finetune Generator | ECG generation · ECG benchmark | FID: 12.26 -> 10.46 | Helped |
| Use TGDP instead of Finetune Generator | ECG synthetic-data augmentation · ECG downstream classification | AUC: 0.941(04) -> 0.953(05) | Helped |
| Use TGDP instead of Finetune Generator | ECG synthetic-data augmentation · ECG downstream classification | Fβ=2: 0.761(10) -> 0.773(11) | Helped |
| Use TGDP instead of Finetune Generator | ECG synthetic-data augmentation · ECG downstream classification | Gβ=2: 0.528(12) -> 0.534(11) | Helped |
| Add cycle regularization to the guidance objective (ablation) | TGDP guidance network · Not specified | performance | Helped |
| Add consistency regularization to the guidance objective (ablation) | TGDP guidance network · Not specified | performance | Helped |
| Use consistency regularization alone instead of the combined guidance objective (ablation) | TGDP guidance network · Target distribution with limited data | performance | Hurt |
Surprising findings
- Vanilla domain-classifier guidance (using the classifier's log(1-c) term) does not reproduce the target distribution; the density-ratio form is required.
- Direct finetuning on very limited target data (only 10 samples) is outperformed by the guidance approach, and consistency-only guidance training is insufficient.
Limitations
- No empirical validation on language or vision tasks, which the authors leave to future work.
- Homogeneous transfer with unseen target classes relies on an unspecified coupling between source and target labels.
Experimental setup
Datasets: PTB-XL ICBEB2018 Gaussian mixture simulation
Benchmarks: ICBEB2018 (downstream classification) Gaussian mixture simulation
Metrics: average likelihood FID coverage (diversity) AUC F_beta=2 G_beta=2
Baselines: Vanilla Diffusion Finetune Diffusion Finetune Generator Vanilla Classifier Finetune Classifier
Builds on: score-based SDE diffusion denoising score matching DPM-Solver classifier guidance
Closest papers by meaning
- Domain Guidance: A Simple Transfer Approach for a Pre-trained Diffusion Model (ICLR 2025 · similarity 0.93)
- Elucidating the design space of classifier-guided diffusion generation (ICLR 2024 · similarity 0.90)
- No Training, No Problem: Rethinking Classifier-Free Guidance for Diffusion Models (ICLR 2025 · similarity 0.90)
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 2025 · similarity 0.90)
- Plug-and-Play Diffusion Distillation (CVPR 2024 · similarity 0.89)
- TFG: Unified Training-Free Guidance for Diffusion Models (NeurIPS 2024 · similarity 0.89)
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.