🤖 machine learning

Assessing Sample Quality in Conditional Generation under Compositional Shift

This paper introduces a post-hoc, per-sample trust score that evaluates the quality of conditional generations under compositional shift by combining global realism and attribute-wise faithfulness using only the training distribution, thereby enabling effective filtering and ranking of extrapolated samples without requiring a reference target distribution.

Original authors: Berker Demirel, Valentino Maiorca, Marco Fumero, Theofanis Karaletsos, Francesco Locatello

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Berker Demirel, Valentino Maiorca, Marco Fumero, Theofanis Karaletsos, Francesco Locatello

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 have a super-smart artist (a computer program called a Conditional Generator) who can paint pictures based on your descriptions. You can ask for "a blue cat," "a smiling dog," or even "a cell with a specific genetic change."

Usually, this artist is great at painting things it has seen before. But what happens when you ask for something brand new? Maybe you ask for a "purple cat with a striped tail," a combination the artist has never seen in its training photos.

In the world of science, this is a huge deal. Scientists want to use these artists to simulate experiments they haven't done yet (like testing a new drug on a cell) because doing the real experiment is expensive or takes too long. But there's a big problem: How do you know if the artist's new painting is actually good and accurate, when you don't have a real photo of that exact thing to compare it to?

This is the "circular problem" the paper addresses. You can't check the new painting against a real photo if the real photo doesn't exist yet.

The Solution: The "Trust Score"

The authors created a new tool called a Trust Score. Think of it as a quality inspector that doesn't need a "perfect reference photo" to do its job. Instead, it uses two simple rules to judge every single painting the artist makes:

1. The "Realism" Check (Does it look like a real thing?)

Imagine the artist has a huge gallery of real photos it learned from. The inspector checks: "Does this new painting look like it belongs in that gallery?"

If the artist paints a cat with six legs or a face made of soup, the inspector says, "Nope, that doesn't look like anything real."
Metaphor: It's like a food critic tasting a new dish. Even if they've never tasted this exact recipe, they know if the texture and ingredients look like a real meal or if it's just colored mud.

2. The "Faithfulness" Check (Did it listen to your instructions?)

This is the tricky part. The inspector asks: "Did the artist actually paint what you asked for, or did it just paint something that looks real but is wrong?"

If you asked for a "striped tail" but the artist painted a "spotted tail," the inspector needs to catch that.
The Trick: Since the inspector has never seen a "striped tail" on a "purple cat" before, it looks at the artist's past work. It asks: "In all the times the artist painted 'striped tails' on other animals, did they look like this? And in all the times it painted 'spotted tails', did they look like this?"
If the new painting is much closer to the "striped" examples than the "spotted" ones, the inspector gives it a pass.
Metaphor: Imagine you ask a chef for a "spicy chocolate cake." You've never had one before. The taste tester doesn't have a "perfect spicy chocolate cake" to compare it to. Instead, they check: "Does this taste more like the spicy cakes we've made before, or the sweet ones?" If it leans toward the spicy side, they trust the chef listened.

The "Magic" Part: Checking While It's Painting

Usually, you have to wait until the painting is 100% finished to check it. But this paper found a way to check the painting while it's still being created.

The Analogy: Imagine the artist is painting a picture on a canvas. Usually, you wait until the paint is dry to judge it. But this new method lets you peek at the canvas halfway through.
By looking at the "skeleton" or "rough draft" of the image while the computer is still thinking, the system can say, "Oh, this is going to be a bad painting," and stop the process immediately.
Why this matters: It saves a massive amount of computer time and energy. You don't waste resources finishing a picture that you know is going to be rejected.

What Did They Find?

The team tested this on two things:

Faces (CelebA): They asked the artist to combine different facial features (like "smiling + blonde hair + glasses") in new ways. The Trust Score successfully filtered out the bad combinations and kept the good ones, even for combinations the artist had never seen before.
Cell Images (RxRx1): This is the scientific part. They asked the artist to simulate cells with specific genetic changes.
- They used a special microscope tool (CellProfiler) to measure the actual shape and texture of the simulated cells.
- The Result: The cells that got a "High Trust Score" from the system looked much more like real biological cells than the ones that got a low score. This means the system can help scientists pick the best "simulated experiments" to test in the real lab, saving time and money.

The Bottom Line

This paper gives us a way to trust AI-generated data even when we are exploring completely new territory. It provides a "lie detector" for AI art and science simulations that works without needing a perfect reference photo, and it can even stop the AI from wasting time on bad ideas before they are fully finished.

In short: It's a quality control system that says, "I don't know what this specific thing should look like, but I know if it looks real, and I know if it listened to your instructions. So, I can tell you if it's worth keeping."

Technical Summary: Assessing Sample Quality in Conditional Generation under Compositional Shift

Problem Statement

Conditional generative models, particularly diffusion models, are increasingly used to simulate conditions that are rare, expensive, or unobserved in real data (e.g., specific biological perturbations or novel attribute combinations). However, evaluating the quality of these "extrapolative" generations presents a fundamental circularity: standard metrics like Fréchet Inception Distance (FID) or Kernel Inception Distance (KID) require a reference distribution of real samples from the target condition to compare against. By definition, this target distribution is unavailable in the extrapolative regime where the model is most useful. Consequently, existing post-hoc reliability tools either measure global realism (compatibility with the training distribution) or attribute alignment (agreement with a prompt) but fail to combine these into a unified, reference-free per-sample trust score that can filter or rank generations when the target ground truth is missing.

Methodology

The authors propose a post-hoc, per-sample trust score ( $T$ ) that assesses conditional sample quality using only the real training distribution and a pretrained feature extractor ( $\Phi$ ). The score decomposes quality into two geometrically estimable quantities:

Global Realism ( $R$ ): Measures the compatibility of a generated sample with the real data manifold. This is computed as a standardized Mahalanobis energy in the feature space of $\Phi$ , quantifying how far the sample is from the empirical mean of real training features.
Attribute-wise Faithfulness ( $F$ ): Measures whether a sample is closer to the requested attribute values than to plausible alternatives. Instead of requiring a real sample for the specific joint condition $a^*$ , the method evaluates each constituent attribute $k$ independently. It computes a margin between the distance to the requested attribute prototype and the minimum distance to competing attribute prototypes. This relies on a reference coverage condition: the observed training support must contain enough variation to isolate attribute effects (e.g., a reference condition where all other attributes are fixed while one varies).

The final trust score is the sum: $T = R + F$ . Lower scores indicate higher trust.

Implementation Variants

Post-generation Scoring: The image is fully decoded, passed through $\Phi$ , and scored.
During-generation Scoring: To enable early rejection, the authors train a lightweight translator (a shallow MLP) to map intermediate denoising representations from the diffusion model directly into the $\Phi$ -compatible feature space. This allows the trust score to be evaluated before full decoding, enabling the rejection of low-quality samples early in the generation trajectory.

Key Contributions

Reference-Free Trust Score: The paper recasts per-sample quality assessment as a problem solvable without observing the joint target conditioning. It introduces a score computable from a pretrained generator, an encoder, and training-set statistics, decomposing quality into global realism and attribute-wise faithfulness.
Attribute-Level Identifiability: The authors theoretically demonstrate that while full conditional fidelity to a missing target distribution is unidentifiable without assumptions, meaningful attribute-level comparison scores are identifiable under a mild reference coverage condition. This formalizes the boundary between what can be certified and what remains fundamentally unidentifiable.
Early Abstention Mechanism: By aligning the diffusion model's internal representations with the feature extractor space via a translator, the method enables quality assessment and sample rejection during the denoising process, avoiding the computational cost of full decoding and re-encoding for low-trust candidates.
Empirical Validation: The score is validated on biological imaging (RxRx1) and controlled vision benchmarks (CelebA), showing it can effectively filter, rank, and curate synthetic data under support shift.

Experimental Results

The method was evaluated on:

RxRx1: A dataset of fluorescent microscopy images with cell types and siRNA perturbations.
CelebA: A compositional testbed with binary face attributes.

Key Findings:

Filtering and Ranking: Under support shift (unseen joint conditions), the trust score consistently selected samples with significantly lower $\Delta$ KID (improvements of 39–44% on CelebA and 39–42% on RxRx1) compared to random baselines. The score also provided a strong Spearman correlation ( $\rho \approx 0.9$ ) between mean trust and $\Delta$ KID, enabling effective condition-level ranking.
Downstream Utility: In RxRx1, samples selected by the trust score improved downstream classification performance (cell type and perturbation identification) when used as training data.
Biological Validity: On RxRx1, trust-selected samples were closer to real morphological centroids in CellProfiler feature space, preserving biologically meaningful structures (texture, shape, granularity) better than unfiltered generations.
During-Generation Efficiency: The translator-based scoring achieved comparable filtering performance to post-generation scoring on CelebA (e.g., +31% $\Delta$ KID improvement) while saving significant compute by rejecting samples early in the denoising trajectory (e.g., saving ~67% of steps at $k=83$ ).
Ablation: The authors showed that standard alignment methods (like REPA's cosine alignment) failed to preserve the necessary covariance geometry for trust scoring on RxRx1, whereas their whitened-geometry translator succeeded.

Significance and Claims

The paper claims to address a critical gap in the deployment of conditional generative models for scientific and exploratory settings where target data is unavailable. By providing a per-sample, reference-free reliability estimate, the method allows researchers to:

Filter unreliable synthetic data before it contaminates downstream analyses.
Rank generated conditions by difficulty or likelihood of success, prioritizing wet-lab experiments.
Abstain from generating low-quality samples early, reducing computational waste.

The authors emphasize that their approach does not improve the generative model itself but provides a necessary diagnostic tool for trust. They note that the method relies on the assumption of reference coverage; if attributes are perfectly confounded in the training data (e.g., only appearing in fixed combinations), attribute-level faithfulness cannot be disentangled. The work is modest in scope, focusing on discrete attributes and demonstrating that while full conditional fidelity is unidentifiable, a decomposition into realism and attribute-wise faithfulness yields a practical, calibrated metric for compositional generalization.