Federico Betti
Research Papers

Hallucination Early Detection in Diffusion Models
Published

HEaD+ (Hallucination Early Detection+) identifies missing-object failures early in the diffusion process using cross-attention, textual information, and a Predicted Final Image signal, enabling early restarts with different seeds to improve completeness while reducing wasted compute.

International Journal of Computer Vision (IJCV) | Published: January 2026
Authors: F Betti, L Baraldi, L Baraldi, N Sebe, R Cucchiara
Diffusion Models Text-to-Image Hallucination Detection Seed Selection Cross-Attention Efficiency Evaluation Generative AI

What Changed? Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models

This paper introduces DICE (DIfference Coherence Estimator), a model designed to detect localized differences between original and edited images and assess their relevance to modification requests using Multimodal Large Language Models.

ICCV 2025 | Published: May 2025
Authors: L Baraldi, D Bucciarelli, F Betti, M Cornia, L Baraldi, N Sebe, R Cucchiara
View PDF
Computer Vision Image Editing Multimodal LLM Evaluation Metrics Human Judgment DICE Instruction-Guided Generative AI Self-Supervision Inpainting

Optimizing Resource Consumption in Diffusion Models through Hallucination Early Detection

We introduce HEaD (Hallucination Early Detection), a novel paradigm that swiftly detects incorrect generations in diffusion models during early stages, saving computational resources and accelerating the generation process for complex multi-object images.

Proceedings of the European Conference on Computer Vision Workshops. 2024. | Published: January 2024
Authors: F Betti, L Baraldi, L Baraldi, R Cucchiara, N Sebe
View PDF
Computer Vision Diffusion Models Resource Optimization Computational Efficiency Hallucination Detection Multi-object Generation Cross-attention Maps Energy Consumption

Let's ViCE! Mimicking Human Cognitive Behavior in Image Generation Evaluation

We introduce ViCE, a novel automated method for evaluating consistency between generated images and text prompts by combining Large Language Models with Visual Question Answering in a process inspired by human cognitive assessment.

Proceedings of the 31st ACM International Conference on Multimedia, 9306-9312 | Published: October 2023
Authors: F Betti, J Staiano, L Baraldi, L Baraldi, R Cucchiara, N Sebe
View PDF
Computer Vision Image Generation Cognitive Science Evaluation Metrics Visual Question Answering LLMs Vision-Language Models

Controlled text generation with adversarial learning

This paper presents a novel architecture for conditioned text generation using adversarial learning, allowing users to guide the generation process through simple sentences or concepts as input.

13th International Conference on Natural Language Generation (INLG 2020) | Published: December 2020
Authors: F Betti, G Ramponi, M Piccardi
View PDF
NLP Text Generation Adversarial Learning GAN Conditioned Generation Deep Learning
© 2026 Federico Betti
Last updated: January 10, 2026