Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Face Consistency Benchmark for GenAI Video

Michal Podstawski, Malgorzata Kudelska, Haohong Wang

TL;DR

The paper addresses the problem of maintaining consistent facial representations for characters in AI-generated videos. It introduces the Face Consistency Benchmark (FCB), which uses widely adopted face embedding models via the DeepFace library to quantify facial similarity and coherence across frames. The evaluation spans four text-to-video models (Runway Gen-3, HunyuanVideo, Vchitect-2.0, CogVideoX1.5-5B) and two comparison modes that measure frame-to-reference similarity and intra-video frame-pair coherence, with real videos serving as a baseline. Findings indicate that while some models show improvements over others, current systems fall short of real-video consistency, underscoring the need for standardized benchmarks to drive progress toward more reliable, character-driven AI video generation.

Abstract

Video generation driven by artificial intelligence has advanced significantly, enabling the creation of dynamic and realistic content. However, maintaining character consistency across video sequences remains a major challenge, with current models struggling to ensure coherence in appearance and attributes. This paper introduces the Face Consistency Benchmark (FCB), a framework for evaluating and comparing the consistency of characters in AI-generated videos. By providing standardized metrics, the benchmark highlights gaps in existing solutions and promotes the development of more reliable approaches. This work represents a crucial step toward improving character consistency in AI video generation technologies.

Face Consistency Benchmark for GenAI Video

TL;DR

The paper addresses the problem of maintaining consistent facial representations for characters in AI-generated videos. It introduces the Face Consistency Benchmark (FCB), which uses widely adopted face embedding models via the DeepFace library to quantify facial similarity and coherence across frames. The evaluation spans four text-to-video models (Runway Gen-3, HunyuanVideo, Vchitect-2.0, CogVideoX1.5-5B) and two comparison modes that measure frame-to-reference similarity and intra-video frame-pair coherence, with real videos serving as a baseline. Findings indicate that while some models show improvements over others, current systems fall short of real-video consistency, underscoring the need for standardized benchmarks to drive progress toward more reliable, character-driven AI video generation.

Abstract

Video generation driven by artificial intelligence has advanced significantly, enabling the creation of dynamic and realistic content. However, maintaining character consistency across video sequences remains a major challenge, with current models struggling to ensure coherence in appearance and attributes. This paper introduces the Face Consistency Benchmark (FCB), a framework for evaluating and comparing the consistency of characters in AI-generated videos. By providing standardized metrics, the benchmark highlights gaps in existing solutions and promotes the development of more reliable approaches. This work represents a crucial step toward improving character consistency in AI video generation technologies.
Paper Structure (5 sections, 1 figure, 2 tables)

This paper contains 5 sections, 1 figure, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Proposed solution
  4. Next steps
  5. Conclusions

Figures (1)

  • Figure 1: Comparison of face consistency in real and AI-generated videos. The evaluation verifies video generation models using similarity between the face in different frames, measured by cosine distance (lower is better). Mode 1 (left) compares all frames to a representative frame. Mode 2 (right) assesses temporal consistency through random frame pairs. Results are averaged over 30 videos.