Research map / Label-efficient and robust learning / Self-supervised and contrastive
Self-Correcting Self-Consuming Loops for Generative Model Training
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 self-correction functions for self-consuming generative model training, proves stability under idealized correction, and validates with a physics-simulator corrector (UHC) on MDM human motion synthesis and a K-means corrector on MNIST.
Problem
Self-consuming loops, where generative models are retrained on their own synthetic outputs, can lead to training instability and model collapse, especially as synthetic data proliferates online.
Approach
Introduce a correction operator that maps synthetic samples toward the true data distribution (idealized as a mixture with the optimal model, strength gamma); prove this makes iterative fine-tuning exponentially more stable, and approximate the corrector at scale using expert knowledge such as a physics simulator.
Key results
- On human motion with 100% synthetic augmentation on 64-example data, self-correction keeps FID lower and more stable across 50 generations, where uncorrected iterative fine-tuning degrades.
- Theory: with correction strength gamma, the stability contraction factor rho(lambda)/(1+gamma) is strictly smaller than the gamma=0 case, giving exponentially better stability.
- On MNIST, self-correction with K=16 K-means centroids improves FID and stability across generations, with larger gains at higher augmentation percentages.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add distribution-wise self-correction, increasing correction strength | Gaussian self-consuming training · Gaussian toy | Wasserstein distance | Helped |
| Add self-correction by mapping generated images to a training-set centroid | MNIST iterative fine-tuning · MNIST | FID | Helped |
| Remove self-correction from iterative fine-tuning (ablation) | MNIST class-conditional diffusion model · MNIST | FID | Hurt |
| Add UHC physics-simulator correction to generated motions | MDM iterative fine-tuning · HumanML3D test (MoVi subset) | FID | Helped |
| Remove UHC self-correction from iterative fine-tuning (ablation) | MDM · HumanML3D test (MoVi subset) | FID | Hurt |
| Replace iterative fine-tuning with ground-truth-only fine-tuning | MDM · HumanML3D test (MoVi subset) | FID | Helped |
| Increase synthetic augmentation percentage (ablation) | MNIST iterative fine-tuning with self-correction · MNIST | FID | Helped |
| Vary the number of centroids used for image correction (ablation) | MNIST iterative fine-tuning · MNIST | FID | No clear effect |
| Use a high synthetic augmentation percentage with UHC self-correction | MDM iterative fine-tuning · HumanML3D test (MoVi subset) | FID | Helped |
| Add any positive correction strength instead of no correction | Theoretical iterative fine-tuning · Self-consuming generative-model training | Stability | Helped |
Surprising findings
- Self-correction allows stable self-consuming training even at a 100% synthetic-to-real ratio, where the theory alone would require small augmentation.
Limitations
- Theoretical guarantees rely on an idealized distribution-wise correction and local concavity/Lipschitz assumptions near the optimum.
- Practical physics-based corrector is only demonstrated on human motion; image experiments use a simple K-means heuristic on MNIST.
Experimental setup
Datasets: HumanML3D MoVi MNIST
Benchmarks: HumanML3D test split (546 sequences) MNIST
Metrics: FID Diversity Matching Score Wasserstein distance (Gaussian toy)
Baselines: iterative fine-tuning without correction baseline fine-tuning on real data only
Builds on: Bertrand et al. self-consuming loop stability analysis Human Motion Diffusion Model (MDM) Universal Humanoid Control (UHC)
Closest papers by meaning
- A Theoretical Perspective: How to Prevent Model Collapse in Self-consuming Training Loops (ICLR 2025 · similarity 0.92)
- On the Stability of Iterative Retraining of Generative Models on their own Data (ICLR 2024 · similarity 0.89)
- Towards Theoretical Understandings of Self-Consuming Generative Models (ICML 2024 · similarity 0.88)
- Self-Consuming Generative Models with Curated Data Provably Optimize Human Preferences (NeurIPS 2024 · similarity 0.87)
- Self-Consuming Generative Models with Adversarially Curated Data (ICML 2025 · similarity 0.87)
- Collapse or Thrive: Perils and Promises of Synthetic Data in a Self-Generating World (ICML 2025 · similarity 0.86)
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.