Research map / Label-efficient and robust learning / Domain adaptation and generalization
Diffusion-Driven Progressive Target Manipulation for Source-Free Domain Adaptation
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 DPTM, a diffusion-based SFDA framework that splits target data into trust and non-trust sets, semantically manipulates non-trust samples toward randomly assigned labels with a latent diffusion model, and progressively refines the pseudo-target domain over iterations.
Problem
Source-free domain adaptation suffers from unreliable pseudo-labels under large domain gaps for non-generation methods, and generation-based methods that create pseudo-source data inherit a source-target shift that limits performance.
Approach
Generate a pseudo-target domain directly (rather than pseudo-source) by manipulating unreliable target samples toward newly assigned labels with a latent diffusion model, using target-guided initialization (FFT low/high frequency mixing), semantic feature injection via DDIM inversion, and domain-specific feature preservation; iterate to shrink the non-trust set progressively.
Key results
- On Office-Home (ResNet-50), DPTM reaches 91.2% average accuracy versus 81.1% for the prior SOTA ProDe.
- On DomainNet-126, DPTM averages 85.2% versus 81.5% for ProDe.
- On VisDA (ResNet-101), DPTM achieves 97.6% per-class accuracy versus 88.7% for ProDe.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply DPTM instead of DM-SFDA | Office-31 · Office-31 (Avg.) | accuracy: 93.7 -> 95.8 | Helped |
| apply DPTM instead of ProDe | Office-Home · Office-Home (Avg.) | accuracy: 81.1 -> 91.2 | Helped |
| apply DPTM instead of DIFO | VisDA · VisDA Sy→Re | accuracy: 88.6 -> 97.6 | Helped |
| apply DPTM instead of ProDe | DomainNet-126 · DomainNet-126 (Avg.) | accuracy: 81.5 -> 85.2 | Helped |
| increase the confidence threshold E from 0.001 to 0.01 (ablation) | DPTM · Office-Home (Avg.) | accuracy: 80.7 -> 91.2 | Helped |
| increase the confidence threshold E from 0.005 to 0.01 (ablation) | DPTM · Office-Home (Avg.) | accuracy: 86.7 -> 91.2 | Helped |
| add progressive refinement, from r=0 to r=10 (ablation) | DPTM · Office-Home (Avg.) | accuracy: 62.2 -> 91.2 | Helped |
| add Target-guided Initialization to Semantic Feature Injection and Domain-specific Feature Preservation (ablation) | DPTM · Office-Home (Avg.) | accuracy: 82.5 -> 91.2 | Helped |
| add Semantic Feature Injection to Target-guided Initialization and Domain-specific Feature Preservation (ablation) | DPTM · Office-Home (Avg.) | accuracy: 79.8 -> 91.2 | Helped |
| add Domain-specific Feature Preservation to Target-guided Initialization and Semantic Feature Injection (ablation) | DPTM · Office-Home (Avg.) | accuracy: 82.2 -> 91.2 | Helped |
| use SDXL instead of Stable Diffusion v1.5 (ablation) | DPTM · Office-Home (Avg.) | accuracy: 75.6 -> 75.6 | No clear effect |
| use the original pseudo-labels instead of uniformly reassigned labels for non-trust samples (ablation) | DPTM · Office-Home first four tasks | accuracy: 86.7 -> 80.3; 94.2 -> 91.1; 92.8 -> 90.1; 91.5 -> 83.9 | Hurt |
Surprising findings
- Randomly reassigning labels to non-trust samples outperformed using their original pseudo-labels for generation (e.g., Ar->Cl 86.7 vs 80.3).
- Setting all target samples to the non-trust set (no trust set) yields little improvement across refinement iterations, with accuracy plateauing near 56% on Ar->Cl.
Limitations
- Gap between synthetic pseudo-target and real target remains, requiring a trust set and BNM alignment to mitigate.
- Performance depends on threshold E and refinement count R; the paper notes other values may do better, and ablations use differing settings (R=3, E=0.001 vs R=10, E=0.01).
Experimental setup
Datasets: Office-31 Office-Home VisDA DomainNet-126 VLCS
Benchmarks: Office-31 Office-Home VisDA DomainNet-126 VLCS
Metrics: classification accuracy (%)
Baselines: CPGA ASOGE ISFDA DM-SFDA SHOT NRC GKD HCL AaD AdaCon CoWA ELR
Builds on: DATUM SHOT BNM DDIM inversion Stable Diffusion
Closest papers by meaning
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.93)
- Dynamic Target Distribution Estimation for Source-Free Open-Set Domain Adaptation (AAAI 2025 · similarity 0.90)
- Invertible Projection and Conditional Alignment for Multi-Source Blended-Target Domain Adaptation (AAAI 2025 · similarity 0.89)
- Understanding and Improving Source-free Domain Adaptation from a Theoretical Perspective (CVPR 2024 · similarity 0.89)
- Vicinity-Guided Discriminative Latent Diffusion for Privacy-Preserving Domain Adaptation (NeurIPS 2025 · similarity 0.89)
- Revisiting the Domain Shift and Sample Uncertainty in Multi-source Active Domain Transfer (CVPR 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.