Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FlexID: Training-Free Flexible Identity Injection via Intent-Aware Modulation for Text-to-Image Generation

Guandong Li, Yijun Ding

TL;DR

FlexID tackles the conflict between identity fidelity and text editability in training-free T2I by decoupling identity into semantic (SIP) and visual (VFA) streams and introducing Context-Aware Adaptive Gating (CAG). SIP injects soft semantic priors into the text space while VFA enforces pixel-level fidelity in the latent space, with CAG dynamically balancing these influences based on editing intent and diffusion timestep. The approach achieves state-of-the-art balance between identity preservation and text adherence on the IBench benchmark without any fine-tuning, demonstrating superior editability in complex narratives. This work offers a practical, plug-and-play framework for controllable multimodal generation and provides a foundation for extending dynamic gating to broader video and multi-subject editing tasks.

Abstract

Personalized text-to-image generation aims to seamlessly integrate specific identities into textual descriptions. However, existing training-free methods often rely on rigid visual feature injection, creating a conflict between identity fidelity and textual adaptability. To address this, we propose FlexID, a novel training-free framework utilizing intent-aware modulation. FlexID orthogonally decouples identity into two dimensions: a Semantic Identity Projector (SIP) that injects high-level priors into the language space, and a Visual Feature Anchor (VFA) that ensures structural fidelity within the latent space. Crucially, we introduce a Context-Aware Adaptive Gating (CAG) mechanism that dynamically modulates the weights of these streams based on editing intent and diffusion timesteps. By automatically relaxing rigid visual constraints when strong editing intent is detected, CAG achieves synergy between identity preservation and semantic variation. Extensive experiments on IBench demonstrate that FlexID achieves a state-of-the-art balance between identity consistency and text adherence, offering an efficient solution for complex narrative generation.

FlexID: Training-Free Flexible Identity Injection via Intent-Aware Modulation for Text-to-Image Generation

TL;DR

FlexID tackles the conflict between identity fidelity and text editability in training-free T2I by decoupling identity into semantic (SIP) and visual (VFA) streams and introducing Context-Aware Adaptive Gating (CAG). SIP injects soft semantic priors into the text space while VFA enforces pixel-level fidelity in the latent space, with CAG dynamically balancing these influences based on editing intent and diffusion timestep. The approach achieves state-of-the-art balance between identity preservation and text adherence on the IBench benchmark without any fine-tuning, demonstrating superior editability in complex narratives. This work offers a practical, plug-and-play framework for controllable multimodal generation and provides a foundation for extending dynamic gating to broader video and multi-subject editing tasks.

Abstract

Personalized text-to-image generation aims to seamlessly integrate specific identities into textual descriptions. However, existing training-free methods often rely on rigid visual feature injection, creating a conflict between identity fidelity and textual adaptability. To address this, we propose FlexID, a novel training-free framework utilizing intent-aware modulation. FlexID orthogonally decouples identity into two dimensions: a Semantic Identity Projector (SIP) that injects high-level priors into the language space, and a Visual Feature Anchor (VFA) that ensures structural fidelity within the latent space. Crucially, we introduce a Context-Aware Adaptive Gating (CAG) mechanism that dynamically modulates the weights of these streams based on editing intent and diffusion timesteps. By automatically relaxing rigid visual constraints when strong editing intent is detected, CAG achieves synergy between identity preservation and semantic variation. Extensive experiments on IBench demonstrate that FlexID achieves a state-of-the-art balance between identity consistency and text adherence, offering an efficient solution for complex narrative generation.
Paper Structure (21 sections, 6 equations, 4 figures, 1 table)

This paper contains 21 sections, 6 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Works
  3. Personalized Text-to-Image Generation
  4. Training-Free Identity Preserving Generation
  5. Cross-Modal Semantic Projection and Feature Disentanglement
  6. Method
  7. Overview
  8. Semantic Identity Projector (SIP)
  9. Visual Feature Anchor (VFA)
  10. Context-Aware Adaptive Gating (CAG)
  11. Semantic-Driven Intent Adaptation
  12. Diffusion-Stage Complementary Scheduling
  13. Final Dynamic Weight Formula
  14. Experiments
  15. Experimental Settings
  16. ...and 6 more sections

Figures (4)

  • Figure 1: Overview of the FlexID Framework. Given a reference image, the identity features are disentangled into a Semantic Identity Projector (SIP) and a Visual Feature Anchor (VFA). The SIP provides semantic-level soft guidance to the text encoder, while the VFA provides visual-level structural constraints to the DiT backbone. The Context-Aware Adaptive Gating (CAG) mechanism dynamically balances the contributions of both streams based on the editing intent of the prompt and the current timestep.
  • Figure 2: Qualitative Comparison. FlexID achieves superior visual atmosphere integration and identity-consistent editability in complex scenes (e.g., stylized oil painting, fisheye lens, atmospheric dining, backlit portrait) compared to other tuning-free methods.
  • Figure 3: Effectiveness of the Semantic Identity Projector (SIP). Left: Without SIP, the identity appears isolated from the environment. Right: With SIP, the identity naturally integrates with the environmental atmosphere and style.
  • Figure 4: Core Role of the Intent-Aware Gating (CAG) Mechanism. Left: Static weighting leads to rigid poses and loss of aesthetic details. Right: CAG enables flexible pose and enriches details while preserving ID.