Source-linked AI summary
EverAnimate: Minute-Scale Human Animation via Latent Flow Restoration
Wuyang Li, Yang Gao, Mariam Hassan, Lan Feng, Wentao Pan, Po-Chien Luan, Alexandre Alahi
TL;DR
Long-form human animation is limited by accumulated quality and identity drift across chunk-wise generation. EverAnimate uses persistent latent memory and restorative flow matching, achieving the best performance across rollout horizons and improving stability at longer horizons.
Problem
Long-form human animation remains constrained by short horizons and accumulated quality and identity drift across chunk-wise generation.
Method
EverAnimate propagates persistent multi-view latent memory across chunks and restores emerging drift through restorative flow matching during sampling.
Results
EverAnimate consistently achieves the best performance across rollout horizons; at 10s, PSNR increases from 23.47 to 25.24 and LPIPS decreases from 0.217 to 0.169 versus Wan-Animate.
Takeaways & Limitations
Together, persistent latent propagation and restorative flow matching enable more stable long-form generation with improved consistency and fidelity.
Abstract
from arXiv · showhide
We propose EverAnimate, an efficient post-training method for long-horizon animated video generation that preserves visual quality and character identity. Long-form animation remains challenging because highly dynamic human motion must be synthesized against relatively static environments, making chunk-based generation prone to accumulated drift: (i) low-level quality drift, such as progressive degradation of static backgrounds, and (ii) high-level semantic drift, such as inconsistent character identity and view-dependent attributes. To address this issue, EverAnimate restores drifted flow trajectories by anchoring generation to a persistent latent context memory, consisting of two complementary mechanisms. (i) Persistent Latent Propagation maintains a context memory across chunks to propagate identity and motion in latent space while mitigating temporal forgetting. (ii) Restorative Flow Matching introduces an implicit restoration objective during sampling through velocity adjustment, improving within-chunk fidelity. With only lightweight LoRA tuning, EverAnimate outperforms state-of-the-art long-animation methods in both short- and long-horizon settings: at 10 seconds, it improves PSNR/SSIM by 8%/7% and reduces LPIPS/FID by 22%/11%; at 90 seconds, the gains increase to 15%/15% and 32%/27%, respectively.
1 Introduction
Long-form human animation is limited by accumulated low-level quality and high-level identity drift during chunk-wise generation, especially across minute-scale horizons. EverAnimate addresses both through persistent latent context memory and intrinsic flow restoration during sampling.
- Limitations: Existing human animation methods typically generate only clips of a few seconds, while autoregressive chunk-wise methods remain limited to hundreds of frames.These approaches struggle to extend animation length substantially.
- Failure modes: Repeated cross-chunk conditioning causes low-level quality drift, while semantic inconsistency produces high-level identity drift.The heterogeneous motion of rapidly changing humans and comparatively stable backgrounds makes long-form generation vulnerable to both drift types.
- Failure modes: Repeated latent-to-pixel reconstruction damages visual details, whereas limited semantic memory cannot reliably anchor long horizons.Attention sinks provide a positive signal but are insufficient for preserving identity over extended generation.
- EverAnimate: EverAnimate is an efficient post-training framework that preserves visual quality and character identity in minute-scale long animation videos.It introduces implicit flow restoration during latent flow propagation, anchored by persistent context memory.
- EverAnimate: Persistent Latent Propagation maintains cross-chunk semantic continuity through latent continuation and multi-view identity memory.The mechanism propagates semantic memory autoregressively across generated chunks.
- EverAnimate: Restorative Flow Matching improves within-chunk visual fidelity by encouraging drift correction during sampling.Together, the two components propagate semantic memory while correcting emerging drift intrinsically.
2 Related Work
Prior human-animation methods progressed from video-to-video translation and explicit motion transfer to pose-conditioned image-to-video generation, improving visual fidelity and controllability. Video foundation models extend temporal context through spatiotemporal compression and scaling, but long-horizon autoregressive continuation remains vulnerable to train-test mismatch, exposure bias, accumulated errors, and forgetting.
- Early methods used video-to-video translation and explicit motion-transfer formulations to model articulated dynamics.
- Recent approaches extract intermediate motion representations, such as 2D poses, before animating reference images with image-to-video generation.Examples include AnimateDiff, MagicAnimate, and Animate Anyone, alongside variants that strengthen pose conditioning and temporal consistency.
- Video foundation models extend effective temporal context through stronger spatiotemporal compression and larger model and data scales.
- Autoregressive extrapolation beyond the training horizon suffers from train-test mismatch, exposure bias, accumulated errors, and forgetting.These limitations motivate complementary work on long-horizon continuation and drift control.
3 Preliminaries and Motivation
This section formulates long-video pose-controlled animation and identifies two sources of accumulated drift in chunk-wise generation: repeated VAE round-trips and insufficient attention-sink anchoring. It motivates persistent latent memory and active trajectory restoration as requirements for stable long-form animation.
- Problem Setup: The task generates a video from a reference image and pose-control sequence while preserving the reference identity and appearance.Pose maps control target motion, while video VAE encoding and decoding operate in latent space.
- Problem Setup: Long-video generation divides the pose sequence into consecutive chunks, generating the first from the reference image and later chunks using decoded, re-encoded carry-over frames.The described carry-over case uses one frame and can extend to sliding windows.
- Problem Analysis: Attention sinks alone cannot fully prevent semantic and visual drift: Wan-Animate still exhibits noticeable long-horizon drift despite most tokens attending to the persistent reference frame.The limitation persists even though the reference image globally anchors identity and appearance across chunks.
- Problem Analysis: Repeated frame-level VAE round-trips inevitably accumulate drift, even when the DiT adds no error on temporally static regions and predicts identical residuals across chunks.The standard pipeline repeatedly decodes, extracts the last frame, and re-encodes it for the next chunk.
- Remark: Stable long-form animation requires latent-space propagation of motion, identity, and appearance semantics plus active sampling-time drift correction.The proposed direction maintains short-term motion and long-term identity memory and uses ODE-based trajectory restoration.
4 Method
EverAnimate combines persistent latent memory with restorative flow matching to generate long animations chunk by chunk while mitigating high-level identity drift and low-level trajectory drift. Its context memory propagates motion and identity across chunks, while velocity adjustment trains recovery from perturbed in-chunk states.
- Method overview: EverAnimate trains on adjacent chunks, extracting motion and identity memories from V (1) to generate V (2) with restorative flow matching.At test time, it rolls out chunk by chunk in latent space without decoding frames between chunks.
- Persistent Latent Propagation: Persistent Latent Propagation propagates motion memory for short-term continuity and identity memory for multi-view consistency across chunks.Motion memory retains the last r latent slices, while identity memory uses sampled frames from the context chunk.
- Persistent Latent Propagation: Random multi-view sampling lets users provide arbitrary reference views while reducing systematic view-to-view bias in identity memory.The method also applies mild identity-preserving spatial augmentation, such as random translation and rescaling, to address context bias.
- Restorative Flow Matching: Restorative Flow Matching trains the denoising flow to recover from nearby perturbed states while preserving the same memory and control pathways used during inference.The model perturbs in-chunk states rather than transmitted context and learns a constant velocity toward the clean endpoint over the remaining interval.
- Restorative Flow Matching: A bounded bell-shaped time weight emphasizes restoration in intermediate trajectory regions and reduces correction near high-noise and low-noise extremes.When ξ = 0, the restorative formulation reduces to standard flow matching; when ξ ≠ 0, it retains a bounded pull toward the clean endpoint.
5 Experiments
Experiments evaluate EverAnimate across multiple long-video durations using frame-level, perceptual, and identity-focused metrics. EverAnimate maintains background and identity quality over long rollouts, achieves the best quantitative performance across horizons, and benefits from both restorative flow matching and persistent latent propagation.
- Datasets and Metrics: Evaluations use 10s, 30s, 60s, and 90s targets at 25 FPS, measuring frame fidelity, perceptual similarity, and identity consistency.Training combines Champ, UBC, Seedance, and 2k self-collected 480P minute-scale YouTube videos.
- Qualitative Comparison: EverAnimate maintains stable background quality and human identity across rollout horizons, avoiding the gradual deterioration and obvious artifacts seen in most models.The qualitative comparison finds most models plausible at short horizons but increasingly degraded over time.
- Quantitative Comparison: At 10s, PSNR increases from 23.47 to 25.24 and LPIPS decreases from 0.217 to 0.169 over Wan-Animate.The quantitative comparison reports EverAnimate as the best-performing method across rollout horizons, while noting that camera-motion differences can reduce PSNR and SSIM reliability.
- Ablation Study: At 60s, the full model improves over the baseline by +5.39 PSNR and increases SSIM from 0.543 to 0.855.Removing RFM lowers perceptual quality, while removing PLP weakens cross-chunk carry-over and long-horizon consistency.
6 Conclusion
EverAnimate is a lightweight post-training framework for long-form pose-guided human animation that addresses long-horizon degradation through latent-state control. Persistent latent memory and restorative flow correction improve generation stability, consistency, and fidelity.
- Conclusion: EverAnimate addresses long-horizon degradation through latent-state control rather than image-space continuation.It is designed for long-form pose-guided human animation.
- Conclusion: Persistent Latent Propagation maintains reusable latent memory across chunks to preserve identity and motion cues over extended rollouts.The mechanism targets temporal consistency across chunked generation.
- Conclusion: Restorative Flow Matching applies a bounded corrective step during sampling to steer perturbed latent trajectories toward the clean generation path.This mechanism complements persistent latent propagation.
- Conclusion: Together, the two designs enable more stable long-form generation and substantially improve consistency and fidelity.The conclusion attributes these improvements to their combined operation.