社内勉強会_Gloria: Consistent Character Video Generation via Content Anchors

Gloria: Consistent Character Video Generation via Content Anchors CVPR 2026

Background ◆ Digital characters High-quality static character images Long-form production Short clips generation Image generation → Motion and temporal coherence in short clips → Extended narratives and reusable digital characters Generative AI is moving from isolated visual content to scalable, character-driven storytelling. ➢ Commercial applications - live streaming - virtual influencer - episodic series

Background ◆ Three key goals of character video generation Long duration Generation Multi-view appearance consistency Expressive identity consistency Generate extended video sequences while maintaining character identity over time Preserve a character’s appearance across different viewpoints and camera angles Maintain the same character identity across different facial expressions and emotions Across time Across viewpoints Across expressions

Motivation ◆ Existing problems Insufficient Identity Context Non-character-centric Memory Limited reference images cannot fully capture a character’s appearance across different viewpoints and facial expressions. Historical video frames contain redundant motion and scene information, rather than explicitly preserving stable character identity. Both limitations lead to identity drift and inconsistent character appearance. How to provide stable and comprehensive character information throughout long-form video generation?

Proposal ◆ Character-centric content anchors - represent character visual attributes using a compact set of anchor frames - provide persistent visual references to maintain consistency throughout long-form video generation Global anchor (Cg) Viewpoint anchors (Cv) Expression anchors (Ce) Preserve global scene context and subject–background relationship Capture character appearance from different viewpoints (front / back / left / right…) Capture character identity across different facial expressions. (neutral / happy / angry / sad...) Long-term consistency Multi-view consistency Expressive identity consistency

Proposal ◆ Challenges of reference-based video generation 1. Copy-pasting 2. Multi-reference Conflicts The model may directly reproduce reference frames instead of learning transferable character attributes. Different anchors may interfere with one another, making it difficult to distinguish and use the right reference information. How to use multiple content anchors without copying or interference?

Method ◆ Mechanism 1: SCA Challenge 1: Copy-pasting Training clips Superset content anchoring (SCA) Content anchors Sample anchors from inside and outside the target training clip to discourage direct duplication. Whole video SCA: expand anchor sampling to learn transferable character features

Method ◆ Mechanism 2: RoPE Challenge 2: Multi-reference conflicts Rotary Position Embedding (RoPE) Assign distinct positional offsets to anchors so the model can distinguish their roles. Video Tokens → normal positions Global Anchor → offset 600 Viewpoint Anchors → offset 200 Expression Anchors → offset 400

Method ◆ Overall pipeline 1. 2. 3. 4. Anchor sampling & Encoding: sample character-centric anchors and encode them into latent space Anchor differentiation: assign distinct RoPE positional offsets to distinguish anchor types and subcategories Multimodal conditioning: integrate content anchors with text, image, and audio conditions through attention mechanisms Decoding & Video generation: denoise video latents and decode them into video frames

Experiments ◆ Setup ➢ Dataset • Large-scale single-person video clips • Training data covering diverse character appearances, motions, and expressions ➢ Evaluation metrics • Long-term: subject consistency, background consistency, ArcFace • Multi-view: DINO-I, CLIP-I • Expressive: Exp.

Experiments ◆ Results: long-term consistency Method Subject ↑ Background ↑ ArcFace ↑ HunyuanAvatar 0.914 0.909 0.495 WanS2V 0.872 0.864 0.439 InfiniteTalk 0.952 0.942 0.695 Gloria 0.960 0.951 0.787 Gloria achieves the best long-term subject, background, and facial identity consistency.

Experiments ◆ Visualization results

Experiments ◆ Results: multi-view and expressive ID Method DINO-I ↑ CLIP-I ↑ Exp. ↑ Phantom 0.597 0.785 0.645 Vidu 0.662 0.822 0.644 Kling 0.698 0.797 0.599 Humo 0.810 0.816 0.672 Gloria 0.821 0.858 0.717 Gloria outperforms competing methods in appearance similarity and expressive identity consistency.

Experiments ◆ Ablation study Long-term Consistency Setting Subject ↑ Background ↑ ArcFace ↑ Full Gloria 0.960 0.951 0.787 w/o Global Anchor 0.840 0.844 0.233 w/o SCA 0.922 0.899 0.489 Setting DINO-I ↑ CLIP-I ↑ Exp. ↑ Full Gloria 0.821 0.858 0.717 w/o Viewpoint / Expression Anchors 0.801 0.848 0.668 w/o RoPE 0.817 0.855 0.671 Appearance & Expression Consistency

Summary Limitations Goal Method Long-form character consistency Character-centric content anchors Limited Generalization Maintain character identity across long video sequences, different viewpoints, and facial expressions. Use global, viewpoint, and expression anchors to provide stable visual references. Primarily focuses on singleperson videos, with limited support for multi-character interactions. SCA reduces copy-pasting, while RoPE helps distinguish different anchors. Requires diverse reference frames and substantial training resources.

Thank you! Q&A