Unified Character Video Editing for Live Streaming

Paper GitHub Hugging Face BibTeX

Zhiyuan Li^1,3Chi-Man Pun^1,*Peng-Tao Jiang^2,*Bo Li²Xiaodong Cun^3,†

¹ University of Macau · ² vivo BlueImage Lab · ³ GVC Lab, Great Bay University

Real-Time
14.47 FPS
Causal Streaming
Instruction-Guided
Long-Stream Consistency
Appearance-Motion Decoupled

Gallery

Edit instruction

Add red frame sunglasses to his face.

Source Video

EditaLive (Ours)

LiveEdit

SANA-Streaming

StreamDiffusionV2

0:000:05

Abstract

EditaLive is a framework for real-time streaming character video editing. Starting from a pretrained image animation model that naturally decouples appearance from motion, it is repurposed for instruction-based human-centric video editing through reference-frame editing and video reconstruction on CharEdit-50K. The model is then adapted from offline bidirectional to causal streaming generation and compressed into a two-step sampler with aligned self-rollout distillation. Fixed RoPE and Align Forcing reduce training-inference discrepancies, while First-Frame Preserved Sparse Attention filters redundant historical information to mitigate appearance drift. Extensive experiments demonstrate state-of-the-art editing performance, faithful facial-expression preservation, and low-latency real-time streaming inference.

Comparison of general video-to-video editing, a cascaded editor-animation pipeline, and the appearance-motion decoupled EditaLive paradigm — Comparison of general video-to-video editing, a cascaded image-editing-and-animation pipeline, and EditaLive's appearance-motion decoupled formulation.

Method

EditaLive addresses streaming character video editing through three consecutive stages. It first reformulates character video editing as appearance editing conditioned on explicit motion signals, then adapts the offline bidirectional video model to causal streaming generation using cached historical context, and finally compresses the causal model into a two-step sampler with aligned self-rollout distillation. Together, these stages provide explicit motion preservation and stable long-term generation.

Three-stage EditaLive pipeline: appearance-motion decoupled editing, causal streaming adaptation, and aligned self-rollout distillation — Overview of the three-stage training pipeline: appearance-motion decoupled editing, causal streaming adaptation, and aligned self-rollout distillation.

EditaLive reformulates character video editing by decoupling the appearance to be edited from the motion to be preserved. A reference frame represents character appearance, while a skeleton sequence and implicit facial representations capture body motion and facial dynamics. For reconstruction-based training, only the reference image is synthetically edited; the real video remains the reconstruction target, paired with a reverse instruction for recovering the original appearance.

Cross-Character Editing

Combine the appearance of a reference character with motion signals extracted from a different driving video.

Example 1

Turn it into a C4D felt doll style.

Driving video

EditaLive

Example 2

Add red frame sunglasses to his face.

Driving video

EditaLive

Comparison

On CharEdit-Bench-S, EditaLive ranks first or second across all eight quality metrics. Character consistency is evaluated with ID-SIM, AED, and APD; editing quality with TA, EQ, BC, and SR; overall video quality with Pick Score; and efficiency with end-to-end FPS and average inter-chunk latency.

Single H100 @ 384 × 67214.47 FPS

Inter-chunk latency0.829 s

Short-video success0.720

Flash-VAED enabled16.4 FPS

CharEdit-Bench-S

Short-video evaluation

150 five-second videos, 81 frames each, at 480 × 832.

Selected quantitative comparisons on CharEdit-Bench-S
Method	Character Consistency			VLM Evaluation				Video Quality	Efficiency
Method	ID-SIM ↑	AED ↓	APD ↓	TA ↑	EQ ↑	BC ↑	SR ↑	Pick Score ↑	FPS ↑	Latency ↓
LiveEdit*	-	-	-	1.542	1.949	1.640	0.393	19.26	17.25	0.696
SANA-Streaming	0.379	0.629	0.381	2.200	1.777	1.582	0.428	19.46	31.20	0.772
StreamDiffusionV2	0.060	0.885	0.485	1.316	1.869	0.473	0.027	19.92	20.96	0.191
LucyEdit*	-	-	-	1.604	1.797	2.121	0.326	18.92	3.083	26.27
UniVideo	0.592	0.605	0.447	2.377	2.182	1.866	0.533	19.54	0.119	680.9
Ditto	0.334	0.685	0.355	1.560	1.796	1.127	0.167	19.49	0.294	275.4
Qwen. + Wan.	0.491	0.541	0.132	2.792	2.544	1.897	0.713	19.55	1.274	63.55
EditaLive	0.550	0.499	0.124	2.796	2.609	2.024	0.720	19.61	14.47	0.829

APD values are multiplied by 10. TA, EQ, BC, and SR denote text alignment, edit quality, background consistency, and success rate. Pick Score evaluates overall video quality. * LiveEdit and LucyEdit do not support global style editing, so their character-consistency metrics are omitted. Efficiency is measured on a single NVIDIA H100 using 81-frame clips at 384 × 672.

CharEdit-Bench-L

Long-video evaluation

30 videos longer than one minute, paired with two prompts for 60 cases.

Selected quantitative comparisons on CharEdit-Bench-L
Method	Character Consistency			VLM Evaluation				Video Quality	Temporal Consistency
Method	ID-SIM ↑	AED ↓	APD ↓	TA ↑	EQ ↑	BC ↑	SR ↑	Pick Score ↑	CLIP ↑	DINO ↑
LiveEdit*	-	-	-	1.583	1.700	1.044	0.017	18.91	97.36	98.38
SANA-Streaming	0.233	0.725	0.402	2.272	1.938	1.121	0.233	19.62	97.28	98.25
StreamDiffusionV2	0.049	0.876	0.329	1.431	2.064	0.638	0.100	19.70	98.42	99.17
Qwen. + Wan.	0.398	0.659	0.126	2.733	2.304	1.776	0.750	19.69	98.30	98.65
EditaLive	0.492	0.576	0.109	2.761	2.678	2.183	0.817	19.78	98.51	98.96

APD values are multiplied by 10. TA, EQ, BC, and SR denote text alignment, edit quality, background consistency, and success rate. Pick Score evaluates overall video quality. EditaLive obtains the strongest overall performance in the long-video setting, including the best TA, EQ, BC, SR, and Pick Score.

Ablation Study

Replacing Align Forcing with Self Forcing causes the largest performance degradation because it introduces a mismatch in KV-cache construction between rollout training and streaming inference. Removing Fixed RoPE exposes the model to unseen positional offsets and weakens reference conditioning, while FPSA filters redundant historical information and preserves first-frame features to mitigate appearance drift.

CharEdit-Bench-L

Aligned self-rollout distillation

Ablation study on aligned self-rollout distillation
Variant	Character Consistency			VLM Evaluation				Video Quality	Temporal Consistency
Variant	ID-SIM ↑	AED ↓	APD ↓	TA ↑	EQ ↑	BC ↑	SR ↑	Pick Score ↑	CLIP ↑	DINO ↑
w/o Align Forcing	0.252	0.711	0.338	2.703	2.567	1.306	0.367	20.03	98.62	99.07
w/o Fixed RoPE	0.381	0.618	0.120	2.728	2.639	1.872	0.617	19.91	98.50	98.95
w/o FPSA	0.455	0.609	0.116	2.737	2.600	2.105	0.800	19.85	98.46	98.88
w/o First-Frame Preservation	0.466	0.604	0.114	2.725	2.622	2.139	0.833	19.84	98.42	98.89
Full EditaLive	0.492	0.576	0.109	2.761	2.678	2.183	0.817	19.78	98.51	98.96

Citation

If you find EditaLive useful for your research, welcome to cite our work using the following BibTeX:

BibTeXarXiv:2608.27123

@article{li2026editalive,
  title   = {EditaLive! Unified Character Video Editing for Live Streaming},
  author  = {Li, Zhiyuan and Pun, Chi-Man and Jiang, Peng-Tao and Li, Bo and Cun, Xiaodong},
  journal = {arXiv preprint arXiv:2608.27123},
  year    = {2026}
}