Crimson AI NewsA CrimsonLingua Network service
EN ع
← Back to news
Research paper Hugging Face

Energy-Guided Flow Matching: Coarse-to-Fine Generation with Moving Endpoints

AI By Crimson AI Hugging Face Papers 19 August 2026 · 00:00 4 views
Share: X Telegram

A new method called Energy-Guided Flow Matching (EG-FM) improves image generation by explicitly modeling a coarse-to-fine trajectory, achieving state-of-the-art FID scores on ImageNet with fewer training epochs.

Energy-Guided Flow Matching: Coarse-to-Fine Generation with Moving Endpoints

Key points

Researchers have introduced Energy-Guided Flow Matching (EG-FM), a novel approach to generative modeling that explicitly models a coarse-to-fine trajectory for image synthesis. The method addresses a key challenge in pixel-space generative models: jointly learning global structure and fine-grained details in high-dimensional space.

Traditional flow matching interpolates noise toward a fixed clean-image endpoint, leaving spectral evolution to be learned implicitly. EG-FM replaces this fixed endpoint with a heat-kernel-filtered endpoint that evolves smoothly from a low-frequency image to a clean image. An image-specific energy-guided scheduling controls the release of high-frequency signals, re-targeting the velocity in flow matching.

The framework requires no adaptation of the backbone or training data, adding negligible cost to both training and inference. In experiments, EG-FM consistently achieved lower FID scores on ImageNet class-conditional generation at 256×256 resolution, reaching an FID of 1.55 at 200 epochs and 1.45 at 600 epochs. At 512×512 resolution, it achieved an FID of 1.58 after only 40 high-resolution adaptation epochs.

EG-FM also transferred successfully to text-to-image generation, scoring 0.85 on GenEval and 83.9 on DPG-Bench. The code is available on GitHub.

TaskResolutionEpochsFID
ImageNet class-conditional256×2562001.55
ImageNet class-conditional256×2566001.45
ImageNet class-conditional512×51240 (adaptation)1.58
Text-to-image--GenEval: 0.85, DPG-Bench: 83.9
Source
Hugging Face · Hugging Face Papers
Related news
Research paper
Hugging Face 29 Aug 2026

Hugging Face Audit: 110 of 124 AI Evaluations Fail to Support Their Claims

A new commit-bound census of 124 Inspect Evals units reveals that 110 stop before deterministic inference due to missing historica...

4
Research paper
Hugging Face 29 Aug 2026

Aphanta: New Framework Diagnoses When Image Editing Boosts Multimodal Reasoning

Hugging Face researchers introduce Aphanta, a diagnostic framework that evaluates when image-editing intermediates improve multimo...

5
Research paper
Hugging Face 29 Aug 2026

EditaLive! Enables Real-Time Character Video Editing for Live Streaming

Hugging Face researchers introduce EditaLive, a framework for real-time human-centric video editing in live streams, achieving sta...

4