Research map / Label-efficient and robust learning / Test-time adaptation
Everything to the Synthetic: Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment
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 Synthetic-Domain Alignment (SDA), which fine-tunes the source model on Mix-of-Diffusion synthetic data so it aligns with the synthetic domain of an unconditional diffusion model used for target-data projection, improving diffusion-driven test-time adaptation.
Problem
Diffusion-driven test-time adaptation projects target data into a diffusion model's synthetic domain, but the source model is misaligned with that synthetic domain, which limits accuracy under distribution shift.
Approach
Instead of only adapting inputs, also fine-tune the source model on labeled synthetic data generated by a conditional diffusion model and then aligned via the same unconditional diffusion model (Mix of Diffusion), so model and target data share one synthetic domain and TTA becomes in-domain prediction.
Key results
- On ImageNet-C (severity 5), SDA improves over DDA by 2.5%-2.9% across ResNet-50, Swin-T, ConvNeXt-T, Swin-B and ConvNeXt-B (e.g., ConvNeXt-B 51.9 vs 49.4).
- On LLaVA 1.5-7b with ImageNet-C, SDA reaches 93.6 average accuracy, a 2.4% gain over DDA.
- On the UniTTA data-stream benchmark, SDA achieves 31.6 average accuracy vs 30.3 for DDA, the best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| align the source model to the unconditional diffusion synthetic domain using MoD | DDA with SDA · ImageNet-C | average accuracy: 44.2 -> 47.0 | Helped |
| align the source model to the unconditional diffusion synthetic domain using MoD | DDA with SDA · ImageNet-W | accuracy: 1.4% to 2.3% | Helped |
| use SDA synthetic-domain alignment | DDA with SDA · CIFAR-10-C | average accuracy: 65.3 -> 72.4 | Helped |
| use SDA synthetic-domain alignment | DDA with SDA · PASCAL VOC-C | average mIoU: 38.6 -> 39.8 | Helped |
| use SDA synthetic-domain alignment | DDA with SDA on LLaVA · ImageNet-C | average accuracy: 91.2 -> 93.6 | Helped |
| use SDA instead of DDA across stream balance/imbalance settings | ResNet-50 · UniTTA | average accuracy: 29.3 -> 31.6 | Helped |
| add conditional diffusion synthetic-data generation for fine-tuning (ablation) | DDA · ImageNet-C | accuracy: 29.7 -> 30.4; 40.0 -> 40.1; 44.2 -> 44.6 | Helped |
| add unconditional diffusion data alignment after conditional generation (ablation) | conditional-generation variant · ImageNet-C | accuracy: 30.4 -> 32.5; 40.1 -> 42.5; 44.6 -> 47.0 | Helped |
| vary synthetic fine-tuning set to 1K images (ablation) | SDA · ImageNet-C | accuracy: 29.7 -> 31.9 | Helped |
| vary synthetic fine-tuning set to 10K images (ablation) | SDA · ImageNet-C | accuracy: 29.7 -> 31.9 | Helped |
| vary synthetic fine-tuning set to 100K images (ablation) | SDA · ImageNet-C | accuracy: 29.7 -> 32.6 | Helped |
Surprising findings
- Source-model accuracy on diffusion-adapted synthetic ImageNet drops by over 18% relative to source data at timesteps t>=500, even though synthetic images look visually indistinguishable from source images.
- Fine-tuning with only one image per class (1K synthetic images) already substantially outperforms DDA, suggesting the model mainly learns the synthetic domain rather than class knowledge.
Limitations
- SDA is orthogonal to target-data adaptation and its gains depend on the underlying diffusion data adaptation method (DDA).
- Requires generating a large synthetic dataset (50K images) and fine-tuning the source model per source model.
- Ensembling with the source model is still needed to handle occasionally unrecognizable adapted images.
Experimental setup
Datasets: ImageNet ImageNet-C ImageNet-W CIFAR-10-C PASCAL VOC-C UniTTA benchmark
Benchmarks: ImageNet-C ImageNet-W CIFAR-10-C PASCAL VOC-C UniTTA ImageNet
Metrics: accuracy mIoU
Baselines: DDA DiffPure GDA MEMO UniTTA TENT CoTTA EATA RoTTA
Builds on: DDA DiffPure GDA DiT ADM
Closest papers by meaning
- Test-time Correlation Alignment (ICML 2025 · similarity 0.89)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.89)
- Towards Real-World Test-Time Adaptation: Tri-net Self-Training with Balanced Normalization (AAAI 2024 · similarity 0.89)
- L-TTA: Lightweight Test-Time Adaptation Using a Versatile Stem Layer (NeurIPS 2024 · similarity 0.88)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.88)
- Diffusion-Driven Progressive Target Manipulation for Source-Free Domain Adaptation (NeurIPS 2025 · 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.