Research map / Video / Video generation
ARGUS: Stacked Multi-View Identity Mosaic Injection for Subject-Preserving Video Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ARGUS, a Wan-2.1-based subject-preserving video generator that injects MLLM-selected image/video identity evidence as a 3x3 stacked mosaic in negative-time, read-only memory, trained without cross-pair data.
Problem
Identity-preserving video generation relies on one or a few static reference images, leading to look-alike fidelity, drift under large yaw/occlusion/scale changes, and reference leakage or brittle cross-pair supervision.
Approach
Represent identity as a diffusion-synchronized 3x3 mosaic of image/micro-clip identity evidence, placed at negative temporal positions in Wan's native token space as read-only memory, with counterfactual nuisance augmentation instead of cross-pair data, plus time-annealed attention and an adaptive learned-skip guidance (ASLG).
Key results
- On OpenS2V-Eval Human-Domain, ARGUS reaches 64.38 Total Score, 4.18 above the strongest baseline (Hailuo 60.20).
- FaceSim reaches 71.86 on OpenS2V-Eval Human-Domain, with NexusScore 51.62 and NaturalScore 79.14.
- On HardID-Celeb, ARGUS obtains 76.80 FaceSim and improves YawScore and OccScore by 12.60 and 15.10 points over the strongest baselines.
Surprising findings
- Replacing counterfactual augmentation with 2K cross-pair samples lowers Total Score by 5.76 points, suggesting nuisance randomization can beat raw cross-pair pairing.
- Vanilla Wan 2.1 without identity injection scores 39.21 Total, far below the identity-injected variants.
Limitations
- Relies on proprietary Nano Banana 2 for high-level semantic counterfactual augmentation.
- Evaluation of identity likeness is partly subjective; HardID-Celeb uses public figures and human studies are in appendix.
Experimental setup
Datasets: OpenS2V-Eval
Benchmarks: OpenS2V-Eval Human-Domain HardID-Celeb
Metrics: Total Score FaceSim NexusScore NaturalScore GmeScore Aesthetics Motion CLIPScore YawScore OccScore
Baselines: Vidu-2.0 Pika-2.1 Jimeng Hailuo ConsisID EchoVideo SkyReels-A2 HunyuanCustom MAGREF VACE-14B Phantom-14B Stand-In
Builds on: Wan 2.1 flow matching perturbed-attention guidance classifier-free guidance RoPE
Closest papers by meaning
- DiasR: Dual-Modal Identity-Anchored Sparse Routing for Efficient Multi-Subject Video Generation (ICML 2026 · similarity 0.90)
- FaithfulFaces: Pose-Faithful Facial Identity Preservation for Text-to-Video Generation (NeurIPS 2026 · similarity 0.88)
- PolyVivid: Vivid Multi-Subject Video Generation with Cross-Modal Interaction and Enhancement (NeurIPS 2025 · similarity 0.87)
- Identity-Preserving Text-to-Video Generation by Frequency Decomposition (CVPR 2025 · similarity 0.86)
- Advancing Narrative Long Video Generation via Training-Free Identity-Aware Memory (NeurIPS 2026 · similarity 0.86)
- Illuminating Visual Identity in Universal Multimodal Embeddings (CVPR 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.