Research map / Image generation / Editing and style transfer

StableAnimator: High-Quality Identity-Preserving Human Image Animation

CVPR 2025 · Image generation › Editing and style transfer · approach: Diffusion face generation and editing · 25 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 StableAnimator, an end-to-end ID-preserving video diffusion framework for pose-driven human image animation, using a global content-aware Face Encoder, a distribution-aware ID Adapter, and HJB-equation-based face optimization at inference.

Problem

Pose-driven human image animation with diffusion models suffers from identity inconsistency and facial distortion, and existing fixes rely on post-processing face-swapping tools that degrade video quality.

Approach

Injects ID information into a video diffusion U-Net via a Face Encoder that attends to reference image embeddings and an ID Adapter that aligns face-attention features to image-attention feature statistics (mean/std) before temporal layers; additionally, an HJB-equation-based optimization of the predicted sample toward ArcFace similarity is interleaved with EDM denoising at inference.

Key results

• StableAnimator reaches CSIM 0.831 on TikTok and FVD 140.62, the best among compared methods, and outperforms ControlNeXt by 47.1% in CSIM.
• On Unseen100, StableAnimator outperforms Unianimate by 45.8% in CSIM (0.805 vs 0.347 for Unianimate as reported in Table 1 setting).
• User study on 30 videos shows users prefer StableAnimator over all competitors for motion, appearance, background, and ID alignment in most cases (e.g., 95.6% vs DisCo for motion).

Measured effects reported in the paper

Surprising findings

• Adding IP-Adapter for ID injection improves CSIM but substantially degrades video fidelity (FVD rises from 371.38 to 484.77 in the ablation) and single-frame quality.

Limitations

• Training data (3K internet videos) is not released, limiting reproducibility.
• Face optimization adds inference-time cost via 10 Adam steps through the VAE decoder and ArcFace at each denoising step.

Experimental setup

Datasets: TikTok dataset Unseen100 self-collected 3K internet videos

Benchmarks: TikTok Unseen100

Metrics: L1 PSNR SSIM LPIPS CSIM FVD

Baselines: MRAA DisCo MagicAnimate AnimateAnyone Champ Unianimate MimicMotion ControlNeXt

Builds on: Stable Video Diffusion AnimateAnyone EDM IP-Adapter

Closest papers by meaning

• MagicAnimate: Temporally Consistent Human Image Animation using Diffusion Model (CVPR 2024 · similarity 0.88)
• FaithfulFaces: Pose-Faithful Facial Identity Preservation for Text-to-Video Generation (NeurIPS 2026 · similarity 0.88)
• When StyleGAN Meets Stable Diffusion: a W+ Adapter for Personalized Image Generation (CVPR 2024 · similarity 0.88)
• $\text{ID}^3$: Identity-Preserving-yet-Diversified Diffusion Models for Synthetic Face Recognition (NeurIPS 2024 · similarity 0.87)
• Identity-Preserving Text-to-Video Generation by Frequency Decomposition (CVPR 2025 · similarity 0.86)
• EverAnimate: Minute-Scale Human Animation via Latent Flow Restoration (NeurIPS 2026 · similarity 0.86)