AR-Edit: Training-Free Autoregressive Video Editing without Inversion
Authors: Hovhannes Margaryan, Quentin Bammey, Christian Sandor, Vicky Kalogeiton
Organizations: Team ARAI, Université Paris-Saclay, CNRS, LISN, France · LIX, École Polytechnique, CNRS, IP Paris, France · LTCI, Télécom Paris, IP Paris, France
Editing a video stream in real time, without access to future frames, is essential for interactive applications. Existing training-free video editing methods, however, assume full offline access to the entire sequence, while streaming approaches either rely on additional training or are limited to specific editing techniques. We show that this limitation can be removed by analyzing the self-attention dynamics of DMD-distilled autoregressive (AR) video diffusion models. We uncover a key asymmetry: queries and keys remain stable and aligned with their source counterparts across denoising timesteps, while values diverge and encode the evolving content. Inherited from the non-distilled bidirectional teacher, this asymmetry is strongest at high noise, where few-step streaming models operate. The resulting mechanism is simple: a single forward pass suffices to capture source identity in the value features, which can be reused throughout generation. Starting from pure Gaussian noise and injecting only these source values reconstructs the input with fidelity matching or surpassing inversion-based methods, effectively eliminating inversion. Based on this insight, we introduce AR-Edit, a training-free, inversion-free method that caches source values once and selectively injects them during denoising. Across multiple AR backbones, AR-Edit enables real-time streaming editing with superior source preservation and tonal fidelity at minimal overhead.
Figures & tables
Figure 1: Training-free video editing paradigms. Inversion-based methods [ 60 , 10 ] (left) invert the source to noise and denoise under the target prompt; inversion-free methods [ 31 , 32 ] (middle) directly transport the source into the edit. Both need many passes of a bidirectional, offline backbone. AR-Edit (right) runs on an autoregressive backbone [ 67 , 21 , 40 ] , processing frames causally: it caches source value features in one pass and injects them at denoising pure noise conditioned on the target prompt.
Figure 2: Q and K are timestep-stable; V is not. Cosine similarity of Q , K , and V to their clean ( t=0 ) counterparts at each self-attention block of Reward Forcing [ 40 ] . Q and K retain high similarity to those obtained from the clean source across all timesteps, while V diverges sharply.
Figure 3: Reconstruction Recipes. Each recipe is named by its initial noise (Inv: inverted, Rand: random), V source (Inv: inversion, Fwd: forward pass), and V timestep ( {Vt} : one V per denoising step, Vt : a single V at a fixed t ). Ours ⟨ Rand, Fwd, V0⟩ removes inversion entirely by starting from random noise with V from a clean forward pass at t=0 .
Figure 4: LPIPS [ 70 ] as a function of the number of pre-denoising NFEs. Ours performs similarly or surpasses all other reconstruction recipes with one pre-denoising NFE.
Table 1: Quantitative comparison with the state of the art on the benchmark. Best in bold , second-best underlined , per backbone block. TF measures preservation of the source intensity distribution (higher = fewer artifacts). Peak Mem. reports allocated / reserved GPU memory.
Figure 6
Figure 7: Visual comparison of editing. Red boxes highlight burning artifacts and identity loss. Ours preserves source appearance while applying the target edit without tonal artifacts.
Figure 8: Source vs. Text adherence for two ablation grids. Shape and colour are fixed by S in (a) and by bmin in (b). Our defaults sit at the balanced knee of the source-text trade-off.
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9: QKV asymmetry on CausVid (DMD-distilled, autoregressive). Cosine similarity of Q , K , and V at each timestep t∈{1000,757,522} to their t=0 (clean source) counterparts, averaged across 120 videos.
Figure 10: QKV asymmetry on CausVid-Bidirectional (DMD-distilled, bidirectional). Cosine similarity of Q , K , and V at each timestep t∈{1000,757,522} to their t=0 (clean source) counterparts, averaged across 120 videos. Q and K retain high similarity to the clean anchor while V diverges, matching its autoregressive version.
Figure 11: QKV asymmetry on the non-distilled bidirectional Wan 2.1 1.3B teacher. Cosine similarity of Q , K , and V to their t=0 (clean source) counterparts, averaged over 120 videos.
Figure 12: QKV asymmetry during generation from noise on Reward Forcing. Cosine similarity of Q , K , and V at each timestep t∈{1000,938,833,625} to their t=0 (clean source) counterparts when generating from pure Gaussian noise with the backbone’s native predict-then-renoise sampler, averaged across 120 videos.
Figure 13: QKV asymmetry during editing on Reward Forcing. Cosine similarity of Q , K , and V at each timestep t∈{1000,938,833,625} to their t=0 (clean source) counterparts during the editing of Algorithm 1 , averaged across 120 videos.
Category
Videos
Aesthetic
Motion
Flow res.
% mov.
Dur. (s)
Motion modification
12
5.10
4.33
4.88
61.4
20.34
Global scene transformation
12
5.39
8.74
8.64
76.6
17.82
Object insertion
12
5.19
10.21
7.77
86.9
34.63
Object swapping
12
4.94
7.34
7.33
70.8
10.53
Change attribute: pose
12
5.11
8.48
9.41
88.5
18.15
Change attribute: material
12
5.05
9.53
8.86
77.2
20.68
Appendix
Table 2: Per-category distribution of the Pexels benchmark. All metrics are means over the 12 clips per category. Aesthetic is the LAION Aesthetic V2 score [ 53 ] ; Motion is the mean dense optical-flow magnitude [ 14 ] between frames at 10% and 60% of the clip duration; Flow res. subtracts the mean flow vector, isolating object from camera motion; % mov. is the fraction of pixels with flow residual above 1.5 ; Dur. is the source clip duration in seconds.
Figure 14: Qualitative reconstructions for the recipes of Figure 3 . Top row: input frames. Each following row is one recipe, with the marker of Figure 3 , run at the smallest pre-denoising NFE its definition allows; the fixed- t recipes appear once per extraction timestep t . Reading top to bottom, the source appearance is lost by inversion alone, restored by every V -injection variant, and fades as the extraction timestep t moves away from the clean source. Ours recovers the input faithfully.
Dynamics edits
Attribute edits
Method
Motion
Global
Insertion
Swap
Pose
Material
Style
Background
Content
Color
All
SDEdit [ 42 ]
4.50/4.75
2.00/ 6.17
4.17/7.50
4.00/ 7.50
4.67/ 5.67
3.42/6.67
2.67/ 4.83
2.58/ 7.50
5.25/6.83
4.83/ 8.00
3.81/ 6.54
FlowEdit [ 31 ]
7.67/ 5.00
5.45/ 5.18
7.55 /6.91
6.40/5.80
6.75 /4.83
7.00/ 6.75
6.25/4.50
5.09/6.18
7.08/6.50
8.17 / 8.17
6.77/5.98
DynaEdit [ 32 ]
8.17 /4.83
5.64 /4.73
7.45/7.45
6.91 /6.36
7.75 /5.00
7.09 /6.27
6.75 /4.50
5.67 / 6.25
7.73 /6.55
8.55 / 8.00
7.17 /5.96
⟨ Inv, Inv, {Vt}⟩
5.67/4.83
3.58/4.25
5.92/ 7.83
4.92/ 6.83
5.50/4.25
4.67/6.67
4.17/4.17
3.67/5.83
6.08/ 7.33
5.17/6.50
4.93/5.85
Ours
7.92 / 5.08
5.50 /4.25
7.92 / 7.67
7.00 /6.50
6.75 / 5.33
7.58 / 8.00
6.83 / 5.00
7.33 /5.00
8.00 / 7.58
8.00/6.67
7.28 / 6.11
Appendix
Table 3: Per-category source and text adherence on our benchmark (Reward Forcing backbone, default configurations of Table 1 ). Each cell reports Source/Text adherence, mean over the 12 clips of the category. Best per column and metric in bold , second-best underlined .
Source Preservation
Text Alignment
Method
PSNR ↑
SSIM ×102↑
LPIPS ×103↓
MSE ×104↓
CLIP-BG ×102↑
CLIP-T ×102↑
CLIP-T (edit) ×102↑
NFE ↓
SDEdit [ 42 ]
18.73
71.70
221.4
213.0
84.28
26.73
22.91
5
FlowEdit [ 31 ]
24.00
86.56
97.6
65.8
90.40
26.19
22.05
9
DynaEdit [ 32 ]
25.33
88.24
83.6
51.3
91.41
25.91
21.91
13
⟨ Inv, Inv, {Vt}⟩
20.42
70.30
183.7
146.9
86.88
25.80
21.15
9
Ours
25.71
86.38
76.0
47.2
91.52
26.21
22.65
6
Appendix
Table 4: Quantitative results with classical source-preservation metrics [ 24 ] on our benchmark. All methods run on the Reward Forcing distilled backbone with the default configurations of the main paper. Best in bold , second-best underlined .
Figure 15: Additional qualitative comparisons. Red boxes highlight artifacts in baselines; our method (orange) preserves source appearance. Edited attributes are underlined .
Figure 16: Visual comparison with StreamEdit (4 editing steps) on Self Forcing. Red boxes highlight artifacts; our method (orange) preserves the source appearance while applying the edit. Edited attributes in the prompts are underlined .
Resolution (H × W)
Chunk size 1
Chunk size 3 (default)
240×416
19.9 / 21.3
20.7 / 22.4
480×832 (default)
23.8 / 26.2
27.2 / 30.5
960×1664
39.4 / 46.0
OOM
Appendix
Table 5: Peak memory of our method across resolutions and chunk sizes. Cells report peak memory (allocated / reserved, GB) on a single H100. OOM = out of memory.
Editing Quality
Overall Visual Fidelity
Recipe
Source Adh. ↑
Text Adh. ↑
Video Quality ↑
CLIP Frame ↑
TF ↑
TF u↑
NFE ↓
⟨ Inv, –, – ⟩
2.58
6.88
7.54
0.781
0.257
0.200
9
⟨ Inv, Inv, {Vt}⟩
4.93
5.85
7.93
0.847
0.383
0.323
9
⟨ Inv, Fwd, V0⟩
5.69
5.37
8.01
0.871
0.378
0.319
10
⟨ Inv, Fwd, V625⟩
5.81
5.12
8.05
0.874
0.389
0.340
10
⟨ Inv, Fwd, V833⟩
5.44
5.08
8.07
0.867
0.370
0.311
10
Appendix
Table 6: Editing quality across reconstruction recipes on the benchmark. Best in bold , second-best underlined . Ours is reported without the commit-then-anchor schedule and text boost for a fair comparison.
Editing Quality
Overall Visual Fidelity
Method
Source Adh. ↑
Text Adh. ↑
Video Quality ↑
CLIP Frame ↑
TF ↑
TF u↑
Ours w/o commit-then-anchor, w/o text boost
5.73
5.42
7.97
0.871
0.405
0.342
Ours w/o commit-then-anchor
5.68
6.60
7.98
0.847
0.401
0.336
Ours w/o text boost
7.20
5.17
7.97
0.900
0.449
0.419
Ours
7.28
6.11
7.96
0.885
0.473
0.408
Appendix
Table 7: Component ablation on the benchmark. Starting from the full Ours configuration, we disable each of the two components in turn: the commit-then-anchor schedule and text boost. Best in bold , second-best underlined .
Figure 17: Effect of each component. Without CtA the edit spreads beyond its target: the chair and the plant behind the cat drift, and the ferry changes from the source. Without text boost the edit is only partly applied: the animal keeps the cat’s build and colour, and the sky keeps its blue. Ours swaps the cat for the dog and turns the sky golden while leaving chair, plant and ferry untouched.
Editing Quality
Overall Visual Fidelity
s
Source Adh. ↑
Text Adh. ↑
Video Quality ↑
CLIP Frame ↑
TF ↑
TF u↑
0
7.20
5.17
7.97
0.900
0.449
0.419
2
7.25
5.60
7.96
0.894
0.465
0.427
4
7.28
6.11
7.96
0.885
0.473
0.408
6
7.16
6.21
7.92
0.876
0.433
0.403
8
7.11
6.27
7.88
0.868
0.438
0.401
Appendix
Table 8: Ablation on the weight s of text-boost. We ablate the text-boost weight from 0 (no boost) to 8 . Best in bold , second-best underlined . Our default is highlighted in orange.
Figure 18: Effect of the text-boost weight s . As s grows, the butterfly turns more orange and the street more snow-covered: at s=0 the edit is incomplete, s=4 completes it, and s=8 adds little beyond it.
Hyper-params
Per-step behaviour
Editing Quality
Overall Visual Fidelity
S
SCtA
Step 0
Step 1
Step 2
Step 3
Source Adh. ↑
Text Adh. ↑
Video Quality ↑
CLIP Frame ↑
TF ↑
TF u↑
0
0
masked
masked
masked
masked
5.81
7.47
7.77
0.796
0.273
0.242
0
1
free
masked
masked
masked
5.47
7.36
7.73
0.787
0.270
0.235
0
2
free
free
masked
masked
5.13
7.42
7.63
0.782
0.267
0.236
0
3
free
free
free
masked
4.43
7.37
7.68
0.772
0.261
0.225
1
1
full
masked
masked
masked
6.34
6.94
7.86
0.829
0.344
0.304
Appendix
Table 9: Ablation over the injection step ( S ) and the CtA start step ( SCtA ). For each configuration we also list the per-step behaviour of the denoising steps: full = full V -injection, masked = masked V -injection (CtA anchor), free = no injection. SCtA=4 effectively disables CtA. Best per column in bold , second-best underlined . Our default is highlighted in orange.
Figure 19: Effect of the injection schedule (S,SCtA) . Small S re-renders the scene around the edit (a yellow canopy appears with the lantern); large S reproduces the source and drops the edit (no lantern). The default S=SCtA=2 inserts the lantern and keeps the row of trees.
Editing Quality
Overall Visual Fidelity
bmin
bCtA
Source Adh. ↑
Text Adh. ↑
Video Quality ↑
CLIP Frame ↑
TF ↑
TF u↑
10
5
7.36
5.60
7.80
0.901
0.481
0.463
16
5
7.23
5.68
7.88
0.899
0.482
0.427
16
10
7.40
5.73
7.92
0.897
0.481
0.437
16
15
7.32
5.81
7.92
0.897
0.479
0.428
22
5
7.14
5.92
7.87
0.886
0.446
0.400
Appendix
Table 10: Ablation over the V-injection block bmin and the commit-then-anchor mask depth bCtA . Best per column in bold , second-best underlined . The orange row marks our default.
Figure 20: Effect of the block thresholds bmin and bCtA . Injecting the source V from early blocks on leaves the source nearly unchanged; injecting only in the last blocks strengthens the edit (the comic-book stylisation, the bending branches) at the cost of source appearance. The default bmin=22 , bCtA=19 sits between the two.
Figure 21: A screenshot from the user study set-up.