One Editor, Many EditsA Unified Training-free Framework for Diverse Video Editing

Adheesh Sunil Juvekar, Onkar Kishor Susladkar, Kiet A. Nguyen, Muntasir Wahed, Nabeel Bashir,
Xiaona Zhou, Tianjiao Yu, Vedant Shah, Ismini Lourentzou
PLAN Lab · University of Illinois Urbana-Champaign
PaperarXivCodeComparisons

TL;DR

We introduce EditVid, a training-free framework for diverse video editing, supporting various instruction-guided and reference-guided edits in one framework, including style transfer, attribute modification, object insertion, part-level editing, and subject replacement.

SourceEᴅɪᴛVɪᴅ
Attribute edit

“Change the elephant's color to blue.”

SourceEᴅɪᴛVɪᴅ
Global tone

“Convert the video to black and white.”

SourceEᴅɪᴛVɪᴅ
Environment appearance

“Change the turquoise waves to black.”

SourceEᴅɪᴛVɪᴅ
Style transfer

“Convert the video style to ukiyo-e.”

SourceEᴅɪᴛVɪᴅ
Dynamic phenomenon

“Change the color of the fire to blue.”

SourceEᴅɪᴛVɪᴅ
Attribute edit

“Change the black clothing to red.”

Abstract

Video editing spans diverse editing paradigms, yet achieving high-quality instruction-guided and subject-guided editing within a single unified framework remains challenging. We introduce EditVid, a training-free framework combining sparse causal memory for local coherence, correspondence-based post-attention token injection for long-range identity preservation, and soft latent blending for edit locality. The same framework supports instruction-guided and reference-guided edits, including style transfer, attribute modification, object insertion, part-level editing, and subject replacement. On FiVE, EditVid achieves 78.16 FiVE-Acc, compared with 58.95 for the strongest evaluated training-free baseline, while obtaining competitive results on IVEBench. A user study further shows a 51.8% overall preference for EditVid over 7 competing methods.

Contributions

  • EditVid unifies instruction-guided and reference-guided video editing without training. The framework extends frozen MM-DiT-based image editors to temporally consistent video editing, supporting style transfer, attribute modification, object insertion, part-level editing, and subject replacement within the same pipeline.
  • Sparse causal memory preserves local temporal coherence. Adjacent-frame spatio-temporal attention shares information across nearby frames, while correspondence-based post-attention token injection preserves long-range appearance and identity. Confidence- and cycle-consistent correspondences match each frame to an anchor before matched visual representations are injected after attention.
  • Soft latent blending maintains edit locality. Timestep-dependent preservation weights derived from source–edit trajectory discrepancies retain source content where needed while allowing the requested modification elsewhere in the video.
  • Benchmarks and a user study assess quality across editing paradigms. EditVid achieves 78.16 FiVE-Acc on FiVE, compared with 58.95 for the strongest evaluated training-free baseline, and obtains competitive IVEBench results. A user study reports 51.8% overall preference for EditVid over seven competing methods.

EditVid Method

EᴅɪᴛVɪᴅ operates only on visual latent tokens inside frozen MM-DiT-based image editors. Text and optional reference-image conditioning remain unchanged, preserving each editor's original semantic control.

EᴅɪᴛVɪᴅ architecture: sparse causal memory, correspondence-based post-attention token injection, and soft latent blending within frozen MM-DiT-based image editors
EᴅɪᴛVɪᴅ combines adjacent-frame spatio-temporal attention, correspondence-based post-attention token injection, and soft latent blending in frozen MM-DiT-based image editors.
01

Sparse causal memory

Each frame attends to visual key–value states from the immediately preceding frame. This bounded context gives strong short-range coherence while avoiding unreliable long-range RoPE interactions.

02

Post-attention token injection

Confidence- and cycle-consistent correspondences match each frame to an anchor. Matched visual representations are injected after attention for long-range appearance and identity preservation.

03

Soft latent blending

Timestep-dependent preservation weights are derived from source–edit trajectory discrepancies, retaining source content where needed while allowing the requested edit elsewhere.

Quantitative Results

EᴅɪᴛVɪᴅ is evaluated on fine-grained editing accuracy, traditional preservation metrics, IVEBench, and curated subject-guided and general video-editing sets.

FiVE editing accuracy

Instruction-following accuracy on the FiVE benchmark. Higher scores indicate better agreement with the requested edit.

FiVE editing accuracy
Method Foundation Subject-guided FiVE-YN ↑ FiVE-MC ↑ FiVE-U ↑ FiVE-O ↑ FiVE-Acc ↑
Pyramid-Edit Video No 33.67 54.01 56.36 31.31 43.84
Wan-Edit Video No 41.41 52.53 55.72 38.22 46.97
FlowDirector Video No 49.16 68.74 70.41 47.49 58.95
StreamEdit (SF) Video No 43.09 61.09 64.11 40.07 51.94
DMT* Image No 34.78 62.06 62.98 33.86 48.42
VideoGrain† Image No 30.50 43.97 44.30 30.17 37.23
TokenFlow Image No 19.36 35.51 36.68 18.18 27.43
VidToMe Image No 20.03 33.50 36.20 17.34 26.77
AnyV2V Hybrid Yes 30.62 45.42 48.96 27.09 38.02
EditVid (Ours) Image Yes 69.45 86.87 88.78 67.54 78.16
* Requires additional depth maps. † Requires additional masks.

FiVE fidelity and perceptual metrics

Video quality and fidelity on FiVE, comparing perceptual quality with preservation of the source content.

FiVE fidelity and perceptual metrics
Method Foundation Subject-guided Str. Dist. ↓ PSNR ↑ LPIPS ↓ MSE ↓ SSIM ↑ CLIP source ↑ CLIP edit ↑ NIQE ↓ MEFS ↑
Source Videos — — 0 ∞ 0 0 100 24.59 19.87 6.33 93.76
Pyramid-Edit Video No 28.65 20.84 276.59 95.63 71.72 26.82 20.20 5.48 80.59
Wan-Edit Video No 12.53 25.57 94.61 41.84 82.55 26.39 21.23 6.54 89.43
FlowDirector Video No 21.56 22.74 107.09 65.20 81.68 27.93 22.00 4.98 81.22
StreamEdit (SF) Video No 10.27 28.47 49.84 21.88 87.52 26.95 21.73 4.38 90.93
DMT* Image No 85.95 14.71 404.60 372.78 51.64 26.66 21.44 5.24 82.30
VideoGrain† Image No 12.40 27.05 185.21 25.10 79.13 25.69 20.31 4.08 88.57
TokenFlow Image No 35.62 19.06 263.61 138.65 72.51 26.46 21.15 4.01 89.00
VidToMe Image No 22.37 21.15 263.91 88.75 70.69 26.84 21.05 4.68 90.06
AnyV2V Hybrid Yes 71.36 15.90 348.59 342.97 50.77 24.89 19.72 5.04 60.36
EditVid (Ours) Image Yes 17.10 25.97 218.71 36.37 83.31 28.83 22.40 4.17 85.31
Source videos are the reference. * Requires additional depth maps. † Requires additional masks.

IVEBench and curated editing evaluation

Editing quality on IVEBench and curated subject-guided and general video-editing examples.

IVEBench and curated editing evaluation
Method Training-free Total Quality Compliance Fidelity Subject · PF ↑ Subject · EQ ↑ Subject · BC ↑ General · PF ↑ General · EQ ↑ General · BC ↑
LucyEditDev No 0.6220 0.8164 0.3679 0.6817 2.019 3.731 6.596 4.667 6.292 8.542
Kiwi-Edit-5B No 0.6936 0.8101 0.5182 0.7524 3.115 4.327 6.596 8.000 8.229 9.458
Qwen-Video-Edit No 0.6765 0.7752 0.5256 0.7286 1.865 2.654 7.654 7.812 7.979 8.958
VidToMe Yes 0.6297 0.7826 0.4010 0.7056 4.308 5.942 5.923 3.875 5.938 5.250
WANEdit Yes 0.6072 0.7973 0.2801 0.7441 5.962 6.654 8.942 5.125 6.625 8.375
FlowDirector Yes 0.6264 0.7771 0.3765 0.7255 4.019 5.865 8.058 4.021 5.521 6.875
StreamEdit Yes 0.6229 0.8238 0.2921 0.7530 4.077 6.673 8.288 5.250 7.000 9.792
EditVid Yes 0.6847 0.8114 0.4587 0.7839 7.769 7.846 9.558 9.458 9.000 9.500
PF: prompt following; EQ: edit quality; BC: background consistency.

Qualitative Results

Select an edit to compare synchronized outputs from the source, representative baselines, and EᴅɪᴛVɪᴅ. Click any video to pause or resume the complete row.

Editing instructionChange the elephant's color to blue.
Source
VidToMe
Pyramid-Edit
Wan-Edit
FlowDirector
StreamEdit
EᴅɪᴛVɪᴅ (Ours)Ours

BibTeX

@article{juvekar2026editvid,
  title={One Editor, Many Edits: A Unified Training-Free Framework for Diverse Video Editing},
  author={Adheesh Sunil Juvekar and Onkar Kishor Susladkar and Kiet A. Nguyen and Muntasir Wahed and Nabeel Bashir and Xiaona Zhou and Tianjiao Yu and Vedant Shah and Ismini Lourentzou},
  journal={arXiv preprint arXiv:2609.04190},
  year={2026}
}