Research map / Video / Video generation
EverAnimate: Minute-Scale Human Animation via Latent Flow Restoration
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 EverAnimate, a LoRA-based post-training method for minute-scale pose-driven human animation that propagates memory in latent space across chunks and adds a restorative flow matching objective that corrects in-chunk drift.
Problem
Chunk-wise autoregressive human animation accumulates low-level background quality drift and high-level identity drift over long horizons (tens of seconds to minutes), and existing anti-drift tricks like attention sinks and image-space carry-over are insufficient.
Approach
Carry a persistent latent context memory (short-term motion latents plus multi-view identity latents) across chunks instead of re-encoding decoded frames, and train with a restorative flow matching loss that adds a bounded velocity correction (Gaussian-scheduled) so the model learns to pull perturbed in-chunk latent states back to the clean trajectory, rather than perturbing the conditioning image
Key results
- At 10s vs. Wan-Animate, PSNR rises from 23.47 to 25.24 and LPIPS falls from 0.217 to 0.169.
- At 90s, the method reaches PSNR 22.646 and FID 23.981, versus 19.676 PSNR and 32.636 FID for Wan-Animate.
- 60s ablation: removing RFM drops PSNR from 23.857 to 21.842; removing PLP drops it to 22.317; baseline reaches 18.472.
Surprising findings
- Perturbing the conditioning image (as in SVI, Helios, Matrix-Game 3.0) is argued to corrupt cross-chunk continuity, motivating perturbing the in-chunk state instead.
Limitations
- Errors still appear in some samples at extremely long lengths, especially out-of-distribution appearances, possibly due to LoRA capacity and data diversity.
- No explicit camera viewpoint or trajectory control.
- No repeated runs or statistical significance tests reported.
Experimental setup
Datasets: Champ UBC Seedance YouTube minute-scale videos (self-collected, 2k)
Benchmarks: in-house long-horizon evaluation at 10s, 30s, 60s, 90s (25 FPS)
Metrics: PSNR SSIM LPIPS FID V-MAE F-PSNR
Baselines: One-to-All SCAIL SteadyDancer UniAnimate-DiT Wan-Animate
Builds on: Wan-Animate flow matching Helios One-to-All Animate attention sink
Closest papers by meaning
- MagicAnimate: Temporally Consistent Human Image Animation using Diffusion Model (CVPR 2024 · similarity 0.89)
- Animate Anyone: Consistent and Controllable Image-to-Video Synthesis for Character Animation (CVPR 2024 · similarity 0.86)
- AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning (ICLR 2024 · similarity 0.86)
- StableAnimator: High-Quality Identity-Preserving Human Image Animation (CVPR 2025 · similarity 0.86)
- PAI-Actor: Cinematic Multi-Actor Character Replacement in Dynamic Scenes (NeurIPS 2026 · similarity 0.86)
- Hallo2: Long-Duration and High-Resolution Audio-Driven Portrait Image Animation (ICLR 2025 · 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.