Research map / Video / Video generation

ReGenHuman: Re-Generating Human Appearances for Realistic Full-Body Video Anonymization

NeurIPS 2026 · Video › Video generation · approach: Diffusion-based human image animation · 0 citations (OpenAlex)

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 ReGenHuman, a video anonymizer that regenerates full-body humans with a LoRA-fine-tuned VACE video-to-video diffusion model conditioned on pose, mask, depth and caption, rather than editing input pixels.

Problem

Full-body human video anonymization that is simultaneously privacy-preserving, realistic, temporally consistent, and useful for downstream tasks; prior methods either redact pixels (losing utility) or generate per-frame (losing temporal coherence).

Approach

Regenerate, don't edit: composite 2D pose, SAM3 human mask, and monocular depth into identity-free conditioning videos (StructHuman, StructAll) and fine-tune a frozen Wan2.1-VACE backbone with LoRA so human regions are synthesized from structure and caption only, giving an architectural guarantee that no original human pixels reach the output.

Key results

• StructHuman reduces clothes-changing re-ID mAP to 2.5% (R-1 1.4%) versus 33.7% for DeepPrivacy2 (body) and 85.9% for FADM on CCVID.
• StructAll attains the lowest text similarity (0.0753) and identity similarity 0.0063 among generative methods, while StructHuman leads Overall Consistency (16.36%), Human Anatomy (89.74%) and Human Identity (19.30%).
• On zero-shot HOIGen video QA with Qwen3-VL-8B, StructHuman scores 82.37% and StructAll 81.80% against an 86.40% un-anonymized upper bound, with FADM at 83.33%.

Surprising findings

• Full-body generative anonymization achieved lower re-identification than blackout, suggesting regenerated appearances mislead re-ID attacks beyond removing identity signal.
• Face-only baselines (EgoBlur) slightly exceed the un-anonymized upper bound on some VQA benchmarks, since body pixels stay intact.

Limitations

• Generation takes minutes per clip, not real-time.
• Reuses source pose verbatim, so gait and silhouette can still identify subjects.
• Text prompt reliance and inherited demographic bias from the base video model; utility drops slightly versus upper bound on VQA.

Experimental setup

Datasets: HOIGen-1M MedVideoCap-55K NExT-QA CCVID

Benchmarks: CCVID HOIGen-1M VBench VBench-2.0 NExT-QA MedVideoCap-55K

Metrics: Identity Similarity Text Similarity R-1 R-5 mAP Subject Consistency Background Consistency Motion Smoothness Temporal Flickering Overall Consistency

Baselines: EgoBlur Gaussian blur Pixelation Blackout FAMS DeepPrivacy2 FADM

Builds on: VACE Wan2.1 LoRA SAM3 DWPose

Closest papers by meaning

• OpenHumanVid: A Large-Scale High-Quality Dataset for Enhancing Human-Centric Video Generation (CVPR 2025 · similarity 0.84)
• VLOGGER: Multimodal Diffusion for Embodied Avatar Synthesis (CVPR 2025 · similarity 0.84)
• MultiPly: Reconstruction of Multiple People from Monocular Video in the Wild (CVPR 2024 · similarity 0.84)
• FaithfulFaces: Pose-Faithful Facial Identity Preservation for Text-to-Video Generation (NeurIPS 2026 · similarity 0.84)
• HyperHuman: Hyper-Realistic Human Generation with Latent Structural Diffusion (ICLR 2024 · similarity 0.84)
• CapHuman: Capture Your Moments in Parallel Universes (CVPR 2024 · similarity 0.83)