🤖 machine learning

Exposing Diversity Bias in Deep Generative Models: Statistical Origins and Correction of Diversity Error

This paper reveals that deep generative models systematically underrepresent data diversity compared to real distributions due to the finite-sample bias of entropy-based metrics like Vendi and RKE, and proposes diversity-aware regularization strategies to mitigate this error.

Original authors: Farzan Farnia, Mohammad Jalali, Azim Ospanov

Published 2026-02-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Farzan Farnia, Mohammad Jalali, Azim Ospanov

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

The Big Picture: The "Copycat" Problem

Imagine you hire a brilliant artist to paint a thousand pictures of a bustling city. You want them to capture the vibe of the city: the noise, the colors, the different types of people, the unique architecture.

The artist does a great job. The paintings look realistic. But if you look closely, you notice something strange: every single painting features the exact same three types of cars, the same five faces, and the same angle of the sun. The artist has become a "copycat." They are so focused on making each individual painting look perfect that they've forgotten to capture the chaotic, messy, diverse reality of the whole city.

This is exactly what this paper discovered about modern AI image generators (like DALL-E, Midjourney, or Stable Diffusion). While they create stunning images, they systematically underestimate the diversity of the real world. They are "safe" but not "wild" enough.

The Core Discovery: The "Small Sample" Trap

The researchers asked: Why does the AI do this?

They found the culprit is a statistical quirk called Finite-Sample Bias.

The Analogy of the Ice Cream Shop:
Imagine you run an ice cream shop with 1,000 flavors.

The Real World (The Population): You have a massive bucket containing a drop of every single one of those 1,000 flavors mixed together.
The AI's Training (The Sample): You only give the AI a tiny spoonful of that bucket to taste.

Because the spoonful is small, it's statistically unlikely to contain every flavor. It might miss the "Lava Lamp" flavor or the "Blue Cheese" flavor entirely. The AI learns from this tiny spoonful. It thinks, "Okay, the world only has the flavors in this spoon."

When the AI tries to generate new ice cream, it only creates flavors it saw in that tiny spoon. It never invents the missing flavors because it doesn't know they exist.

The Paper's Insight:
The researchers proved mathematically that as you take more samples (a bigger spoon), your estimate of the total diversity gets better. But because AI models are trained on finite datasets (even large ones), they are always working with a "spoonful" rather than the "whole bucket." This causes them to naturally produce less variety than the real world.

How They Measured It: The "Vendi Score"

How do you measure "variety" in a pile of AI images? You can't just ask a human to count them; that's too slow and subjective.

The paper uses a clever math tool called the Vendi Score (and a cousin called RKE).

The Analogy of the Party:
Imagine you are at a party.

Low Diversity: Everyone is wearing the exact same t-shirt. The "Vendi Score" is low.
High Diversity: Everyone is wearing something completely different. The "Vendi Score" is high.

The researchers used a "super-ear" (a mathematical kernel) to listen to the "music" of the images. They calculated how many "unique notes" were being played.

The Result: When they played the "music" of real photos (like from ImageNet), the score was high (lots of unique notes).
The Result: When they played the "music" of AI-generated photos, the score was significantly lower. The AI was playing a loop of the same few notes over and over.

The Solution: "Shaking the Dice"

If the problem is that the AI is too "safe" and stuck in a small corner of the diversity map, how do we fix it?

The paper suggests two main strategies, which they tested successfully:

1. Training with a "Diversity Penalty"

The Analogy: Imagine a teacher grading a student.

Old Way: "You get an A if your drawing looks exactly like the reference photo."
New Way: "You get an A if your drawing looks like the photo AND if it's different from the other drawings you just made."

The researchers added a "diversity penalty" to the AI's training. If the AI starts making images that look too similar to each other, it gets a "failing grade" (a penalty). This forces the AI to stretch its legs and explore the "missing flavors" in the ice cream bucket.

2. The "Post-Processing" Filter

The Analogy: Imagine you have a bucket of 1,000 AI-generated photos. Some are great, but many are clones of each other.
Instead of throwing them away, you use a special filter to re-weight them. You say, "I'll keep this rare, weird photo, but I'll throw away the 50 copies of that boring photo."

The paper shows that by mathematically re-arranging the probability of which images get shown, you can instantly boost the diversity score without needing to re-train the whole AI from scratch.

Why This Matters

This isn't just about making "worse" or "better" pictures. It's about honesty.

For Artists: If an AI only shows you the "average" version of a style, it limits human creativity.
For Science: If an AI simulates a biological process but misses the rare, diverse variations, it could lead to bad medical conclusions.
For Society: If an AI generates news or historical data, and it "flattens" the diversity of human experience, it creates a distorted view of reality.

The Takeaway

Deep learning models are amazing, but they have a blind spot: they are naturally conservative. They tend to play it safe and repeat patterns they've seen, missing the rare and the unique.

This paper provides the tools to measure that "conservatism" and the mathematical "nudges" to force the AI to be more adventurous, ensuring it captures the full, messy, beautiful diversity of the real world, not just a safe, small slice of it.

1. Problem Statement

Deep generative models (GANs, Diffusion Models, VAEs) have achieved remarkable success in generating high-quality samples. However, a critical but under-explored issue is whether these models faithfully capture the diversity of the underlying data distribution.

The Core Question: Do trained generative models suffer from a systematic "downward diversity bias," producing samples that are less diverse than the true data distribution?
The Gap: Standard evaluation metrics (e.g., FID, KID) focus on sample quality and distributional similarity but often fail to isolate structural biases arising from finite-sample training. Specifically, it is unclear if the observed lack of diversity is an artifact of the model architecture or a fundamental statistical limitation of learning from finite datasets.

2. Methodology

The authors employ a combination of statistical theory, empirical evaluation, and optimization-based correction strategies.

A. Evaluation Metrics (Reference-Free)

To measure diversity without relying on a reference dataset (which introduces its own biases), the paper utilizes entropy-based diversity scores:

Vendi Score: Defined as the exponential of the von Neumann entropy of a normalized kernel similarity matrix. It captures the effective number of distinct modes in a sample set.
RKE (Rényi Kernel Entropy): An order-2 Rényi entropy analogue of the kernel matrix, offering a computationally tractable alternative.

Feature Embeddings: Evaluations are performed on embeddings extracted from pre-trained models (e.g., DINOv2, CLIP) to ensure semantic relevance.

B. Statistical Analysis of Finite-Sample Bias

The paper investigates the root cause of diversity bias through statistical theory:

Classical Entropy Bias: It is well-known in statistics that plug-in estimators of Shannon entropy from finite samples are downward biased (underestimate true entropy).
Extension to Kernel Entropy: The authors prove that this bias extends to the Vendi score. They demonstrate that the expected value of the log-Vendi score (von Neumann entropy of the empirical kernel covariance operator) is monotonically increasing with sample size $n$ .
Implication: Finite training sets inherently under-represent the diversity of the true population distribution. Consequently, models trained to minimize divergence to these empirical distributions inherit this diversity deficit.

C. Correction Strategies

Based on the theoretical finding that diversity can be improved by moving a distribution toward a "high-entropy region," the authors propose three mitigation strategies:

Training-Time Regularization: Augmenting the loss function with an entropy term (e.g., maximizing VNE or minimizing Inverse-RKE) to encourage the generator to explore a wider support.
Inference-Time Guidance: For diffusion models, applying gradients derived from diversity scores (Vendi or RKE) during the reverse-diffusion process to steer samples toward more diverse regions.
Post-Hoc Projection (Reweighting): A geometric projection method that reweights a fixed set of generated samples to satisfy a target diversity constraint (entropy superlevel set) while minimizing the Kernel Distance (MMD) to the original uniform distribution.

3. Key Contributions

Empirical Discovery of Systematic Bias:
- Across multiple benchmarks (ImageNet, FFHQ, MS-COCO, LSUN) and architectures (StyleGAN-XL, LDM, DiT), the authors show that generated samples consistently achieve lower Vendi and RKE scores than real test data, even when sample sizes are matched.
- This confirms a universal downward diversity bias in modern generative models.
Theoretical Proof of Finite-Sample Origin:
- The paper provides a formal proof (Proposition 1) that the expected log-Vendi score increases monotonically with sample size.
- This establishes that the diversity gap is partly a statistical inevitability of training on finite datasets, as the empirical distribution of the training set has lower entropy than the true population.
Demonstration of Training Set Size Impact:
- Experiments training models on subsets of data (e.g., 10% vs. 100% of ImageNet) show a direct correlation: smaller training sets lead to significantly lower diversity scores in the generated output.
Novel Correction Frameworks:
- The authors introduce Entropy-Based Guidance and Post-Hoc Reweighting as principled methods to correct this bias.
- They prove a Projection Principle (Theorem 1): Projecting a learned distribution onto a high-entropy convex set (defined by Vendi/RKE) reduces the discrepancy (MMD/KD) to the true data distribution if the true distribution lies within that set.

4. Key Results

Diversity Gap: Real datasets consistently outperform generated samples in Vendi and RKE scores. For example, on ImageNet, the Vendi score of real data increases significantly with sample size, while generated samples plateau at a lower level.
Effectiveness of Guidance:
- Applying SPARKE (RKE Guidance) or Vendi Guidance to DiT-XL-2 and DDIM models significantly improves diversity scores (Vendi, RKE, Recall, Coverage).
- Crucially, this improvement in diversity does not compromise quality. In fact, FID (Fréchet Distance) and KD (Kernel Distance) scores often improve alongside diversity, suggesting that the guidance helps the model better align with the true data manifold.
Post-Hoc Reweighting:
- Reweighting generated samples to satisfy a Vendi constraint (via quadratic programming) yields higher diversity scores than uniform sampling.
- Table 1 results show that "Projected-Vendi" sampling achieves higher Vendi/RKE scores with only modest increases in KD/FD, effectively reducing the diversity bias without generating new samples.
Training Size Sensitivity: Models trained on 10% of the data show drastically lower diversity scores compared to those trained on the full dataset, validating the finite-sample bias hypothesis.

5. Significance and Impact

Redefining Evaluation: The paper shifts the focus from purely "quality" metrics (FID) to a more holistic view including "diversity," arguing that high fidelity without diversity is a failure mode of generative modeling.
Theoretical Insight: It bridges classical information theory (entropy estimation bias) with modern deep learning, explaining why models fail to capture full diversity (finite sample effects) rather than just attributing it to model capacity.
Practical Solutions: The proposed guidance and reweighting strategies offer immediate, implementable tools for practitioners to improve the diversity of diffusion models and GANs without retraining from scratch.
Future Directions: The work suggests that future generative models should explicitly optimize for entropy/diversity constraints, moving beyond simple likelihood or divergence minimization to ensure robust coverage of the data manifold.

In summary, this paper provides a rigorous statistical explanation for the "mode collapse" or diversity deficit in generative models and offers a mathematically grounded toolkit to mitigate it, demonstrating that diversity and quality can be simultaneously enhanced.