FAIL: Flow Matching Adversarial Imitation Learning for Image Generation

2026-02-12 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition · Depth: Advanced, quick

Summary

The paper introduces Flow Matching Adversarial Imitation Learning (FAIL), a novel framework for post-training flow matching models by aligning their output distribution with a high-quality target, mathematically equivalent to imitation learning. Unlike Supervised Fine-Tuning, FAIL addresses policy drift without requiring costly preference pairs or explicit reward modeling. It minimizes policy-expert divergence through adversarial training, offering two algorithms: FAIL-PD, which uses differentiable ODE solvers for low-variance pathwise gradients, and FAIL-PG, a black-box alternative for discrete or computationally constrained environments. When fine-tuning FLUX with 13,000 demonstrations from Nano Banana pro, FAIL achieves competitive performance on prompt following and aesthetic benchmarks. The framework also generalizes to discrete image and video generation and acts as a robust regularizer against reward hacking in reward-based optimization.

Key takeaway

For research scientists developing or fine-tuning generative models, FAIL offers a robust method to improve model alignment without the overhead of preference pairs or explicit reward models. You should consider integrating FAIL-PD for continuous settings or FAIL-PG for discrete or resource-limited scenarios to enhance performance on prompt following and aesthetic benchmarks, while also leveraging its regularization capabilities to prevent reward hacking.

Key insights

FAIL aligns flow matching models with target distributions via adversarial imitation learning, bypassing explicit rewards.

Principles

• Post-training flow matching equals imitation learning.
• Adversarial training can minimize policy-expert divergence.

Method

FAIL uses adversarial training to minimize policy-expert divergence, with FAIL-PD for pathwise gradients via differentiable ODE solvers and FAIL-PG for black-box or discrete settings.

In practice

• Fine-tune FLUX models with 13,000 demonstrations.
• Apply to discrete image and video generation.
• Mitigate reward hacking in reward-based optimization.

Topics

• Flow Matching
• Adversarial Imitation Learning
• Image Generation
• Video Generation
• Reward Hacking

Code references

• HansPolo113/FAIL

Best for: Computer Vision Engineer, Research Scientist, AI Researcher, AI Scientist, Deep Learning Engineer

Related on AIssential

• Support-Conditioned Flow Matching Is Kernel Smoothing — Cross-attention conditioning in generative models is kernel smoothing, with predictable failure modes and learned corrections.
• Generative Adversarial Networks (GANs) - Explained — GANs use adversarial competition between two neural networks to learn and generate data indistinguishable from real samples.
• Improving Visual Representation Alignment Generation with GRPO — VRPO uses adaptive, reward-guided representation alignment to dynamically optimize diffusion transformer training for better fidelity and speed.
• Geometric Erasure by Contrastive Velocity Matching in Rectified Flows — GEM unifies trajectory-based and teacher-guided erasure for Rectified Flow models to suppress harmful content.
• Flow Matching with Semidiscrete Couplings — Flow matching trains generative models by aligning velocity fields with target data through ODE integration.
• The Reward Was in Your Data All Along: Correcting Flow Matching with Discriminator-Guided RL — DRL corrects flow matching models by using a discriminator in a pretrained space to provide a data-aligned reward, overcoming L2 loss limitations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.