BeyondFacial: Identity-Preserving Personalized Generation Beyond Facial Close-ups

TL;DR

This paper tackles the limitation of identity-preserving personalized generation that overly concentrates on facial close-ups, leading to weak narrativity and semantic inconsistency. It introduces a Dual-Line Inference framework that decouples identity and scene semantics, along with an Identity Adaptive Fusion strategy that fuses ID and semantic information at the noise-prediction stage, and an Identity Aggregation Prepending module to stabilize target identity. The proposed method yields strong identity fidelity and rich scene semantics beyond faces, demonstrated through extensive experiments, user studies, and multiple applications including cinematic-grade character-scene generation, style control, and gender alignment. The results show the approach is plug-and-play, tuning-free, and capable of delivering film-grade, full-body scenes with coherent actions while preserving target identities, significantly expanding IPPG applicability.

Abstract

Identity-Preserving Personalized Generation (IPPG) has advanced film production and artistic creation, yet existing approaches overemphasize facial regions, resulting in outputs dominated by facial close-ups.These methods suffer from weak visual narrativity and poor semantic consistency under complex text prompts, with the core limitation rooted in identity (ID) feature embeddings undermining the semantic expressiveness of generative models. To address these issues, this paper presents an IPPG method that breaks the constraint of facial close-ups, achieving synergistic optimization of identity fidelity and scene semantic creation. Specifically, we design a Dual-Line Inference (DLI) pipeline with identity-semantic separation, resolving the representation conflict between ID and semantics inherent in traditional single-path architectures. Further, we propose an Identity Adaptive Fusion (IdAF) strategy that defers ID-semantic fusion to the noise prediction stage, integrating adaptive attention fusion and noise decision masking to avoid ID embedding interference on semantics without manual masking. Finally, an Identity Aggregation Prepending (IdAP) module is introduced to aggregate ID information and replace random initializations, further enhancing identity preservation. Experimental results validate that our method achieves stable and effective performance in IPPG tasks beyond facial close-ups, enabling efficient generation without manual masking or fine-tuning. As a plug-and-play component, it can be rapidly deployed in existing IPPG frameworks, addressing the over-reliance on facial close-ups, facilitating film-level character-scene creation, and providing richer personalized generation capabilities for related domains.

Figures (16)

• Figure 1: Existing Identity-Preserving Personalized Generation (IPPG) methods over-rely on facial close-ups. Awkward backgrounds, incomplete characters, and truncated semantics hinder visual storytelling. Ours transcends this limitation, enabling holistic scenes and rich, full-fledged character generation beyond close-ups.
• Figure 2: Under the same random seed, ID embedding transforms full-body images into facial close-ups.
• Figure 3: Overview of framework, integrating three key innovations: (1) Dual-Line Inference (DLI) pipeline, separating identity and semantic streams to resolve their representational conflict; (2) Identity Adaptive Fusion (IdAF) strategy, deferring fusion to noise prediction with adaptive attention/masking for interference-free fusion without manual input; (3) Identity Aggregation Prepending (IdAP) module, aggregating ID info and replacing random initializations to enhance preservation. This framework enables high-quality IPPG beyond facial close-ups via identity-semantic harmonization.
• Figure 4: Qualitative Results: Personalized Generation Beyond Facial Close-Ups. Our method faithfully renders both character actions and prompt-specified semantic scenes. Generated characters exhibit natural poses and high identity fidelity to reference images, while scenes are aesthetically consistent and visually compelling. This demonstrates significant potential for film and television production applications.
• Figure 5: Visual Comparison. Other methods predominantly generate facial close-ups—focusing on heads, lacking full-body poses and actions, with scene semantics only implied through head backgrounds. In contrast, our method naturally renders full-body actions and complete scene contexts, while maintaining high identity fidelity and strong visual appeal.