Calibrating Generative Models to Feature Distributions with MMD Finetuning

2026-06-17 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Life Sciences & Biology · Depth: Expert, quick

Summary

Kernel Calibrating Generative Models (kCGM) is a novel method designed to correct distributional miscalibration in generative models, where samples are plausible but deviate from a target set's feature distribution. Unlike direct finetuning, which can overfit and lacks control over feature matching, kCGM minimizes a maximum mean discrepancy (MMD) between generated and target feature distributions. It achieves this using an unbiased score-function estimator, complemented by KL regularization to maintain proximity to the pretrained model. The method was successfully applied to a target set of 174 antibiotics, where kCGM improved feature matching while simultaneously increasing chemical validity, a significant advantage over direct finetuning. Furthermore, kCGM's versatility was demonstrated in protein and DNA generation tasks, showing its capability to adapt autoregressive, continuous-space diffusion, and discrete diffusion models using only feature-level supervision.

Key takeaway

For AI Scientists developing generative models for specific target distributions, kCGM offers a robust alternative to direct finetuning. You should consider implementing kCGM to improve feature matching while preserving sample validity, especially in domains like drug discovery or protein design. This approach mitigates overfitting risks inherent in simpler finetuning methods and allows adaptation of diverse model architectures using only feature-level supervision.

Key insights

kCGM corrects generative model miscalibration by matching feature distributions with MMD and KL regularization, improving validity.

Principles

• Generative models can misalign target feature distributions.
• Direct finetuning risks overfitting and validity loss.
• MMD minimization aligns generated and target features.

Method

kCGM minimizes MMD between generated and target feature distributions using an unbiased score-function estimator, with KL regularization to stay near the pretrained model.

In practice

• Apply kCGM for targeted drug molecule generation.
• Calibrate autoregressive and diffusion models.

Topics

• Generative Models
• MMD Finetuning
• Distributional Calibration
• Drug Discovery
• Diffusion Models
• Protein Generation

Code references

• smithhenryd/cgm

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Stitching and dimensionality effects on large artificially generated volume datasets — Stitching artifacts in large image generation significantly degrade downstream segmentation, undetected by perceptual metrics like FID scores.
• Drifting Preference Optimization for One-Step Generative Models — DrPO enables efficient preference finetuning for one-step generative models using black-box rewards and feature-space updates.
• Latent Process Generator Matching — Latent Process Generator Matching generalizes generative models by learning from time-dependent, projected latent Markov processes.
• The FID Lottery: Quantifying Hidden Randomness in Generative-Model Evaluation — FID scores are highly variable due to training randomness, making single reported numbers unreliable and requiring new evaluation protocols.
• Support-Conditioned Flow Matching Is Kernel Smoothing — Cross-attention conditioning in generative models is kernel smoothing, with predictable failure modes and learned corrections.
• Building Generative Image & Video models at Scale - Sander Dieleman (Veo and Nano Banana) — Diffusion models generate high-quality audiovisual data through iterative denoising of learned compressed latent representations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.