Conventional video editing primarily focuses on scene-level content, whereas live streaming places greater emphasis on the human subject. However, directly applying existing video-editing methods to human-centric live streaming remains challenging, as they may introduce facial-expression inconsistencies and typically depend on multiple offline inference steps, making them unsuitable for real-time interaction. We propose EditaLive, a novel framework for real-time streaming character video editing. In detail, we start from a pretrained image animation model (Wan-Animate), which naturally decouples appearance from motion, and repurpose it as the base model for instruction-based human-centric video editing by reference frame editing and video reconstruction via the collected CharEdit-50K dataset. Besides, we adapt the model from offline bidirectional to causal streaming generation, and design an aligned self-rollout distillation strategy that compresses the model into a two-step sampler, where fixed RoPE and align forcing reduce training--inference discrepancies, and first-frame preserved sparse attention filters redundant historical information to mitigate appearance drift. Extensive experiments demonstrate that EditaLive delivers state-of-the-art editing performance with faithful preservation of facial expressions and low-latency real-time streaming inference.
Figures & tables
Figure 2: Comparison of three character video editing paradigms. (a) General Video-to-Video Editing directly transforms a source video V into an edited video V′ , but often relies on costly synthetic video pairs that may contain facial-expression mismatches and weakened human details. (b) Cascaded image editing and animation first edits the source appearance A into A′ and then animates it using motion M , introducing cascaded latency and potential long-term appearance drift. (c) EditaLive unifies appearance editing and motion-conditioned generation, learning from edited reference images and motion-aligned real videos to preserve facial expressions and body motion while enabling stable, low-latency streaming editing.
Figure 3: Overview of the three-stage training pipeline of EditaLive. (a) Appearance–Motion Decoupled Editing reformulates character video editing as appearance editing with explicit body and facial motion preservation. (b) Causal Streaming Adaptation converts bidirectional modeling into causal generation using clean historical context and causal attention. (c) Aligned Self-Rollout Distillation compresses the model into a two-step sampler, where Align Forcing and Fixed RoPE align training with streaming inference, and FPSA improves long-term appearance stability.
Figure 4: Construction pipeline of CharEdit-50K. (Left) Existing video editing datasets often contain facial expression inconsistencies. (Right) Our pipeline restricts synthetic editing to the reference image while retaining its authentic source video as motion-aligned supervision. (1) High-quality reference frames are selected based on face and hand quality. (2) GPT-5.5 generates instructions covering add–remove, remove–add, change–restore, and stylization, followed by category and color balancing. (3) Qwen-Image-Edit and Nano Banana 2 synthesize candidate images, which are filtered for edit accuracy, target clarity, character consistency, and reverse feasibility. Each retained bidirectional sample contains an edited image, its reverse instruction, and the corresponding source video.
Figure 5: Comparison of Self Forcing and Align Forcing under two-step distillation. (a) Self Forcing couples gradient update with rollout termination, causing an intermediate prediction to be cached when the rollout stops early. This creates a mismatch with inference, which always caches the final prediction, and requires 2.5 NFEs on average. (b) Align Forcing applies gradients only at the sampled step but completes the remaining rollout with stop-gradient. The KV cache is therefore always constructed from the final prediction, matching inference while requiring two NFEs. (c) During streaming inference, each chunk propagates its final KV cache after two-step denoising, with an additional t0 forward pass for the first chunk to initialize the attention sink.
Figure 6: Qualitative comparisons. EditaLive faithfully follows the editing instructions while preserving the source facial expressions, body poses, and scene structure across frames.
Method
#Params
Character Consistency
VLM Evaluation
Video Quality
Efficiency
ID-SIM ↑
AED ↓
APD ↓
TA ↑
EQ ↑
BC ↑
SR ↑
Pick Score ↑
FPS ↑
Latency ↓
LiveEdit*
1.3B
–
–
–
1.542
1.949
1.640
0.393
19.26
17.25
0.696
SANA-Stream.
2B
0.379
0.629
0.381
2.200
1.777
1.582
0.428
19.46
31.20
0.772
Stream.V2
14B
0.060
0.885
0.485
1.316
1.869
0.473
0.027
19.92
20.96
0.191
LucyEdit*
5B
–
–
–
1.604
1.797
2.121
0.326
18.92
3.083
26.27
UniVideo
13B
0.592
0.605
0.447
2.377
2.182
1.866
0.533
19.54
0.119
680.9
Table 1: Quantitative comparisons on CharEdit-Bench-S. Numbers in red and blue indicate the best and the second-best results, respectively. APD multiplied by 10 . TA, EQ, BC and SR denote text alignment, edit quality, background consistency, and success rate, respectively. *LiveEdit and LucyEdit do not support global style editing. We therefore omit their character-consistency metrics, as these scores are not comparable when the requested edit is not successfully performed.
Figure 7: Ablation study on causal streaming adaptation and aligned self-rollout distillation.
Method
Character Consistency
VLM Evaluation
Video Quality
Temporal Consistency
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 FP
0.466
0.604
0.114
2.725
2.622
2.139
0.833
19.84
98.42
98.89
Ours
0.492
0.576
0.109
2.761
2.678
2.183
0.817
19.78
98.51
98.96
Table 2: Ablation study on aligned self-rollout distillation.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Examples from CharEdit-50K.
Figure 9: Examples from CharEdit-Bench.
Figure 10: Qualitative comparisons on CharEdit-Bench-L. EditaLive consistently maintains the requested edit and character appearance while following the source motion.
Method
Character Consistency
VLM Evaluation
Video Quality
Temporal Consistency
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-Stream.
0.233
0.725
0.402
2.272
1.938
1.121
0.233
19.62
97.28
98.25
Stream.V2
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
Ours
0.492
0.576
0.109
2.761
2.678
2.183
0.817
19.78
98.51
98.96
Appendix
Table 3: Quantitative comparisons on CharEdit-Bench-L. Numbers in red and blue indicate the best and the second-best results, respectively. APD multiplied by 10 . *LiveEdit does not support global style editing. We therefore omit its character-consistency metrics, as these scores are not comparable when the requested edit is not successfully performed.
Figure 11: Cross-character editing results.
Figure 12: Limitations of EditaLive. The editing prompt is “ Transform it into a Snoopy style ”.
Figure 13: More comparisons between baselines and our EditaLive (1/4).
Figure 14: More comparisons between baselines and our EditaLive (2/4).
Figure 15: Long-video comparison results (1/2).
Figure 16: More comparisons between baselines and our EditaLive (3/4).
Figure 17: Long-video comparison results (2/2).
Figure 18: More comparisons between baselines and our EditaLive (4/4).
Streaming video editing has made rapid progress, yet practical deployment is still limited by two core issues: maintaining stable backgrounds and non-edited regions over time, and achieving the low latency required for real-time interactive scenarios. Meanwhile, recent streaming video generation methods are mostly developed for synthesis and cannot be directly applied to editing due to the strict preservation requirement and region-specific control. In this work, we present a novel streaming video editing framework that performs causal, frame-by-frame editing with strong content preservation and real-time responsiveness. Our key design is a three-stage distillation pipeline that progressively transfers editing capability from a powerful bidirectional foundation model to an efficient unidirectional streaming editor, enabling stable long-horizon edits without sacrificing visual fidelity. To further support real-time deployment, we introduce an AR-oriented mask cache that reuses region-related computation across frames, substantially reducing redundant processing and accelerating inference. Finally, we establish a dedicated benchmark for streaming video editing. Extensive evaluations demonstrate that our method achieves state-of-the-art visual quality among streaming baselines while drastically boosting inference speed to 12.66 FPS, making it suitable for interactive and augmented reality applications.
Streaming video effect generation is highly desirable for live human-centric applications such as e-commerce streaming, entertainment, and vlogging, yet remains difficult due to the lack of suitable data and deployable editing models. Unlike generic video generation, this task requires real-time video-to-video editing that adds expressive effects while preserving human identity, background content, and temporal consistency. Existing acceleration efforts mainly focus on text-to-video generation, while efficient distillation for video editing remains largely underexplored. In this paper, we present StreamingEffect, a real-time human-centric streaming video effect framework. We adopt an in-context video editing architecture and train a high-quality bidirectional teacher, then distill it into a causal autoregressive student and further reduce sampling from 50 steps to 4 steps. We also introduce keyframe control, allowing reference effect frames to be injected online and propagated through the stream for interactive editing. To address the data bottleneck, we construct VideoEffect-130K, to our knowledge the largest human-centric video effect dataset, containing 70K effect videos and 60K editing videos across 600 effect categories curated from short-video and editing platforms. Experiments show that our method enables real-time, high-quality 720p video editing on a single H200 GPU.
Video editing aims to modify input videos according to user intent. Recently, end-to-end training methods have garnered widespread attention, constructing paired video editing data through video generation or editing models. However, compared to image editing, the high annotation costs of video data severely constrain the scale, quality, and task diversity of video editing datasets when relying on video generative models or manual annotation. To bridge this gap, we propose LIVE, a joint training framework that leverages large-scale, high-quality image editing data alongside video datasets to bolster editing capabilities. To mitigate the domain discrepancy between static images and dynamic videos, we introduce a frame-wise token noise strategy, which treats the latents of specific frames as reasoning tokens, leveraging large pretrained video generative models to create plausible temporal transformations. Moreover, through cleaning public datasets and constructing an automated data pipeline, we adopt a two-stage training strategy to anneal video editing capabilities. Furthermore, we curate a comprehensive evaluation benchmark encompassing over 60 challenging tasks that are prevalent in image editing but scarce in existing video datasets. Extensive comparative and ablation experiments demonstrate that our method achieves state-of-the-art performance. The source code will be publicly available.
Weicheng Wang, Zhicheng Zhang, Zhongqi Zhang +6
Nankai University · Kuaishou Technology · Pengcheng Laboratory +1