InvDiff: Invariant Guidance for Bias Mitigation in Diffusion Models
Source
- Raw Markdown: paper_invdiff-2025.md
- PDF: paper_invdiff-2025.pdf
- Preprint: arXiv 2412.08480
- Official code: Hundredl/InvDiff
Status And Credibility
InvDiff was posted to arXiv on 2024-12-11. The TimeCraft README labels it as a KDD 2025 paper. It is included in this ingest because TimeCraft references it as a general diffusion technique. The main paper is text-to-image, but it includes a narrow time-series forecasting experiment.
Core Claim
InvDiff mitigates unknown biases in pretrained text-to-image diffusion models without auxiliary bias annotations. It infers latent bias environments through a max-min objective and trains a lightweight module to guide sampling toward invariant semantic information.
Time-Series Relevance
InvDiff is mostly adjacent rather than direct TSFM evidence. Its useful idea for time-series generation is invariant guidance: a generator may need to preserve task-relevant semantics while reducing spurious correlations. For time series, that could translate to reducing shortcut dependence on cohort, season, dataset identity, or prompt artifacts. The paper includes a limited OOD probabilistic time-series forecasting experiment on AusElec using TimeGrad and CRPS; treat that as narrow domain-shift and bias-mitigation evidence for forecasting, not as broad time-series generation or TSFM evidence.
Evidence And Results
The paper evaluates Waterbirds, CelebA, and FairFace image-generation settings with bias, CLIP-T, FID, and recall metrics. It reports bias reduction while maintaining generation quality, with a small trainable module rather than full diffusion-model retraining. The AusElec time-series experiment is much narrower and evaluates OOD probabilistic forecasting with TimeGrad.
Limitations
- InvDiff is primarily a text-to-image diffusion paper, not a time-series generation paper.
- The bias metrics and datasets are visual/social-attribute settings; direct transfer to temporal data is unproven.
- The AusElec experiment uses month/holiday/domain shifts, not actions, control inputs, or interventions.
- Invariant guidance can hide failure modes if the chosen invariant signal is incomplete or misaligned.
Foundation TSFM Relevance
| Agenda slot | Verdict | Evidence | Missing pieces |
|---|---|---|---|
| Benchmark hygiene and bias | adjacent | Provides a diffusion-side invariant-guidance pattern for reducing unknown biases, plus a narrow AusElec/TimeGrad OOD forecasting experiment. | Needs broader time-series tests for dataset identity, cohort shift, seasonality shortcuts, and rare regimes. |
| Time-series generation | insufficient evidence | The numeric experiment is OOD probabilistic forecasting, not generation of synthetic time-series corpora. | Must be adapted and evaluated on temporal generation before counting as TSFM generation evidence. |