🤖 machine learning

From Fake to Real: Pretraining on Balanced Synthetic Images to Prevent Spurious Correlations in Image Recognition

The paper proposes "From Fake to Real" (FFR), a two-step training pipeline that pre-trains models on balanced synthetic data to learn robust subgroup representations before fine-tuning on real data, thereby preventing spurious correlations caused by mixing real and synthetic distributions and significantly improving worst-group accuracy.

Original authors: Maan Qraitem, Kate Saenko, Bryan A. Plummer

Published 2026-08-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Maan Qraitem, Kate Saenko, Bryan A. Plummer

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to recognize animals. You show it thousands of photos, but there's a sneaky problem: almost every picture of a "Big Dog" happens to be taken inside a cozy living room, while almost every picture of a "Small Dog" is taken outside in a park. The robot is smart, but it's also a bit efficient. Instead of learning what a dog actually looks like, it learns a shortcut: "If I see a couch, it's a Big Dog. If I see grass, it's a Small Dog." In the world of computer science, this is called a "spurious correlation." The robot has learned to rely on the background instead of the subject. This is a big deal because when the robot meets a Big Dog outside, it gets confused and fails. Scientists are trying to fix this so that AI can see the world clearly, not just the patterns it was tricked into seeing.

To fix this, researchers have tried a clever trick: they use "generative AI" (like a digital artist) to create new, fake photos to balance the books. If there aren't enough pictures of Big Dogs outside, the computer draws some. The idea was that by mixing these new fake photos with the real ones, the robot would finally learn to ignore the background. But, as it turns out, this mix-and-match approach had a hidden flaw. The new paper, "From Fake to Real," suggests that simply mixing real and fake photos together actually creates a new kind of reliance. The robot starts noticing that the fake photos look slightly different from the real ones—maybe the fake dogs have a weird texture or the lighting is just a tiny bit off. So, the robot learns a new, even sneakier shortcut: "If the photo looks like a real photo, it's a Small Dog. If it looks like a fake drawing, it's a Big Dog." It's not looking at the dog at all; it's looking at whether the image is real or synthetic.

The authors of this paper, Maan Qraitem, Kate Saenko, and Bryan A. Plummer, realized that trying to teach the robot with both real and fake photos at the same time was the problem. They proposed a new two-step training method called "From Fake to Real" (FFR). Think of it like teaching a student for a math test. First, you give them a stack of practice problems that are perfectly balanced and easy to understand, so they learn the core concepts without any confusion. This is the "Fake" stage: the robot trains only on the perfectly balanced, computer-generated images. It learns what a dog looks like without any messy background distractions. Once the robot has a solid foundation, you move to the second step: the "Real" stage. Now, you show it the messy, real-world photos. Because the robot already knows what to look for from the first step, it doesn't get tricked by the background or the fact that the images are real. It just applies what it learned.

The researchers tested this idea on three different datasets, including photos of faces and animals, with bias levels ranging from moderate to extremely severe (up to 99.9% of one group being in a specific setting). They found that their two-step method was a game-changer. While previous methods that mixed real and fake data together struggled as the bias got worse, the "From Fake to Real" approach stayed strong. In fact, it improved the accuracy of the robot on the hardest cases (the "worst group") by up to 20% compared to the best existing methods. When they looked closely at what the robot was actually paying attention to, they saw that the old methods were still distracted by the background or the "fake-ness" of the images, but the new method focused exactly on the dog or the person's face.

The paper also ran some simulations to prove why the old way didn't work. They showed that as long as you mix real and synthetic data, the robot will almost always find a way to use the difference between them as a distinguishing factor. By separating the training into two distinct stages, they effectively blocked the robot from ever seeing the "difference" between real and fake as a clue. The authors note that this method works well even when combined with other techniques, but they also point out a limitation: it relies on the computer artist being able to draw a perfectly balanced set of fake images. If the artist gets the balance wrong in the first step, the whole system suffers. However, for now, this "Fake then Real" strategy offers a surprisingly simple and effective way to stop AI from learning the wrong lessons, ensuring that when it looks at a dog, it sees a dog, not the room it's sitting in.

Technical Summary: From Fake to Real (FFR)

Problem Statement
Visual recognition models are prone to learning spurious correlations when training sets are biased, where a specific condition $B$ (e.g., "Indoors") is over-represented within a specific class $Y$ (e.g., "Big Dogs"). While synthetic data generated by large-scale generative models (e.g., diffusion models) offers a promising avenue to mitigate this by augmenting underrepresented subgroups, existing approaches introduce a new source of bias.

Prior methods, such as Additive Synthetic Balancing (ASB) and Uniform Synthetic Balancing (USB), mix real and synthetic data during training. The authors argue that this creates a confounding variable $G$ (Data Source: Real vs. Synthetic). Because synthetic and real data often possess distributional differences (e.g., generative artifacts), models trained on mixed data may learn to predict $Y$ based on the pair $(B, G)$ rather than the true semantic features. For instance, a model might learn that "Synthetic Women" smile while "Real Women" do not, simply because the synthetic generation process created that correlation. The paper provides a theoretical proof (Theorem 1) demonstrating that in standard spurious-correlation settings, any augmentation of a biased dataset with synthetic data inevitably results in a bias toward the pair $(B, G)$ .

Methodology: From Fake to Real (FFR)
To address the bias toward the data source $(B, G)$ , the authors propose From Fake to Real (FFR), a two-stage training pipeline that separates the learning of synthetic and real data distributions.

Stage 1: Pretraining on Balanced Synthetic Data
The model is pre-trained exclusively on a balanced synthetic dataset ( $D_{syn}$ ) where the target label $Y$ is statistically independent of the bias group $B$ ( $P_{D_{syn}}(Y|B) = P_{D_{syn}}(Y)$ ). This is achieved by sampling an equal number of synthetic images for each bias subgroup using a generative model (e.g., Stable Diffusion). The goal is to learn robust, unbiased representations across all subgroups without exposure to the statistical differences between real and synthetic data.
Stage 2: Fine-tuning on Real Data
The model initialized from Stage 1 is fine-tuned on the biased real dataset ( $D$ ). To prevent the model from relearning the bias present in the real data, this stage utilizes standard Empirical Risk Minimization (ERM) or common loss-based bias mitigation methods (e.g., GroupDRO, Resampling, Deep Feature Reweighting). Crucially, because the model was pre-trained on balanced data, it starts with a representation that is less susceptible to the spurious correlations in the real data.

By separating the data sources into distinct training stages, FFR prevents the model from ever observing the joint distribution of $(B, G)$ , thereby avoiding the learning of spurious signals related to whether an image is real or synthetic.

Key Contributions

Novel Pipeline: Introduction of FFR, a simple, two-stage training framework that utilizes synthetic data for bias mitigation without exposing the model to the distributional gap between real and synthetic data.
Theoretical Analysis: A formal proof showing that mixing real and synthetic data in a single training phase inevitably introduces bias toward the data source $(B, G)$ , rendering prior mixed-training approaches suboptimal.
Comprehensive Evaluation: Systematic analysis across three datasets (CelebA-HQ, UTK-Face, SpuCo Animals) and five bias ratios (ranging from 90% to 99.9%), totaling over 5,000 experiments.
Integration: Demonstration that FFR is compatible with and enhances existing synthetic-data-free bias mitigation methods (e.g., GroupDRO, DFR).

Experimental Results
The authors evaluate FFR against baselines including no synthetic data, USB, and ASB, using Worst Accuracy (WA) and Balanced Accuracy (BA) as metrics.

Performance Gains: FFR improves worst-group accuracy over state-of-the-art methods by up to 20% across the evaluated datasets. For example, on the UTK-Face dataset with a 95% bias ratio, FFR outperforms USB and ASB by over 10%.
Robustness: Unlike prior methods whose performance degrades significantly as bias severity increases (particularly on CelebA-HQ and UTK-Face), FFR maintains stable performance across all bias ratios.
Feature Learning: Qualitative analysis using RISE saliency maps shows that models trained with FFR focus on relevant semantic features (e.g., the dog's face) rather than spurious background features or generative artifacts (e.g., extra toes in synthetic images). In contrast, ASB and USB models often rely on these artifacts.
Embedding Analysis: t-SNE projections reveal that FFR projects real and synthetic data embeddings into a single, tight cluster, indicating the model learns shared core features. Conversely, ASB and USB project real and synthetic data into separate clusters, confirming they rely on source-specific features.

Significance and Limitations
The paper claims that FFR effectively resolves the issue of bias introduced by the distributional differences between real and synthetic data, a problem overlooked by previous synthetic augmentation strategies. By decoupling the training of these two data sources, FFR allows synthetic data to serve as a source of unbiased representations that improve performance on real data.

The authors acknowledge limitations regarding the quality of the generative models used. If the generative model itself is biased or fails to follow prompts faithfully (e.g., struggling with compositionality), the pretraining distribution may remain imbalanced, which significantly hurts FFR performance (as shown in ablation studies). Additionally, the method relies on the generative model's ability to produce a balanced synthetic dataset; if the model cannot generate specific subgroups accurately, the benefits of the first stage are diminished. The authors suggest that future work focusing on training fairer and more accurate generative models would further alleviate these issues.

Technical Summary: From Fake to Real (FFR)

More like this