DynaEdit: Inversion-Free Video Editing

Updated 4 July 2026

DynaEdit is a training-free video editing method that uses pretrained text-to-video flow models to enable complex dynamic modifications including changes in actions, interactions, and global effects.
It employs Similarity Guided Aggregation (SGA) and Annealed Noise Correlation (ANC) to preserve low-frequency structures and suppress high-frequency jitter for coherent results.
DynaEdit outperforms traditional inversion-based and inversion-free techniques by balancing extensive dynamic edits with faithful preservation of scene layout, identity, and camera motion.

DynaEdit is a training-free editing method that unlocks versatile video editing capabilities with pretrained text-to-video flow models, designed for text-guided editing of real videos across content, actions, dynamics, and interactions (Kulikov et al., 18 Mar 2026). In the formulation described for the method, the inputs are a real source video $x^{\text{src}}$ , a source text prompt $c^{\text{src}}$ , a target text prompt $c^{\text{tar}}$ , and optionally an edited first frame $f^{\text{tar}}$ ; the target output $x^{\text{edit}}$ is expected to obey the target text, preserve scene layout, identity, camera motion, and unaffected objects, remain temporally coherent, and maintain photorealistic or stylistically coherent appearance (Kulikov et al., 18 Mar 2026). The method is positioned as training-free, inversion-free, and model-agnostic, and is built on pretrained text-to-video rectified flow models, especially image-to-video variants (Kulikov et al., 18 Mar 2026).

1. Problem domain and conceptual position

DynaEdit addresses a specific gap in controlled video generation: editing real videos in ways that alter not only appearance but also actions, dynamic events, and causal interactions. The motivating examples include changing what an object does, inserting an object that should affect the behavior of other objects, and introducing global effects such as storms, night-time lighting, earthquakes, or lava-like dynamics. The difficulty is structural. To change actions and outcomes, the model must alter coarse spatio-temporal structure; to preserve the source video, it must remain conservative about scene geometry, identity, and unaffected motion (Kulikov et al., 18 Mar 2026).

The method is framed against two existing families of training-free video editing. Inversion-based methods attempt to reconstruct the source video by inverting it into a latent noise representation and then regenerating under a new prompt. In the characterization given for DynaEdit, these methods suffer from imperfect reconstruction, misalignment between source and target trajectories, and frequent dependence on model-specific interventions. Inversion-free methods, such as FlowEdit, avoid source inversion and are training-free and model-agnostic, but are described as effectively limited to structure-preserving edits; when pushed to larger changes, they yield either weak edits or severe artifacts such as misaligned motion, jitter, and blur (Kulikov et al., 18 Mar 2026).

Within that landscape, DynaEdit is presented as an advanced inversion-free method that aims to extend pretrained text-to-video flow models from appearance-level edits to action modification, object insertion with true interaction, object swapping with altered outcomes, and global dynamic effects. A central question attached to the method is whether a pretrained text-to-video flow model can be used as a “world model” and steered at inference time to produce deep dynamic edits while still preserving the source video where appropriate (Kulikov et al., 18 Mar 2026).

2. Rectified-flow formulation and inversion-free editing

DynaEdit is built on the rectified-flow formulation used by text-to-video flow models. The basic object is a velocity field $V$ satisfying the ODE

$dZ_t = V(Z_t, t)\, dt, \quad t \in [0,1].$

In the rectified setting described for the method, the time-marginal distribution is a linear interpolant,

$X_t = (1 - t) X_0 + t X_1,$

where $X_0$ is a data sample and $X_1$ is Gaussian noise (Kulikov et al., 18 Mar 2026). For image-to-video models, the velocity field is written as $c^{\text{src}}$ 0, where $c^{\text{src}}$ 1 is a noisy video at time $c^{\text{src}}$ 2, $c^{\text{src}}$ 3 is the text prompt, and $c^{\text{src}}$ 4 is the conditioning first frame (Kulikov et al., 18 Mar 2026).

The inversion-free editing setup used as DynaEdit’s starting point defines source- and target-conditioned velocities

$c^{\text{src}}$ 5

and an edit-direction field

$c^{\text{src}}$ 6

The editing trajectory itself is defined by

$c^{\text{src}}$ 7

with noisy source and target states

$c^{\text{src}}$ 8

$c^{\text{src}}$ 9

and initialization $c^{\text{tar}}$ 0 (Kulikov et al., 18 Mar 2026). In the discrete version, at each timestep $c^{\text{tar}}$ 1, one samples $c^{\text{tar}}$ 2 noises, computes per-sample edit velocities, averages them, and updates the edit state by

$c^{\text{tar}}$ 3

The preexisting inversion-free framework also introduces a parameter $c^{\text{tar}}$ 4, which controls how much noise is added and therefore mediates a preservation-versus-editability trade-off. Lower $c^{\text{tar}}$ 5 gives stronger structure preservation and weaker edits; higher $c^{\text{tar}}$ 6 gives stronger edits and more artifacts (Kulikov et al., 18 Mar 2026). DynaEdit departs from that trade-off by always using $c^{\text{tar}}$ 7, i.e. starting from the full-noise end, and then compensating for the associated instabilities through new mechanisms (Kulikov et al., 18 Mar 2026).

3. Failure modes of naive unconstrained editing

DynaEdit’s technical contribution is motivated by two failure modes that appear when inversion-free editing is naively extended to unconstrained dynamic edits. The first is low-frequency misalignment. At full noise level, $c^{\text{tar}}$ 8 and $c^{\text{tar}}$ 9 are essentially pure noise, so the earliest edit steps, which dominate coarse motion, layout, and camera trajectory, are determined largely by the noise seed rather than by the source video (Kulikov et al., 18 Mar 2026). The examples described for the method show that changing only the initial noise can alter train speed, camera motion, collision timing, and spatio-temporal alignment between source and edited video. This is the sense in which large-scale structure becomes random relative to the source (Kulikov et al., 18 Mar 2026).

The second failure mode is high-frequency jitter. If independent noise is sampled at each timestep, the edit direction fluctuates across steps, especially in newly created or heavily edited regions. The reported consequence is fuzzy or blurry objects, temporal flicker, and unstable object contours (Kulikov et al., 18 Mar 2026). Using the same noise across all timesteps reduces jitter, but that extreme leads to poorer low-frequency alignment and physically implausible outcomes. The paper’s contrast between independent and fully correlated noise is therefore not merely a variance issue; it exposes a structural tension between global control and local temporal stability (Kulikov et al., 18 Mar 2026).

These two pathologies motivate DynaEdit’s two main mechanisms. Similarity Guided Aggregation (SGA) is introduced to preserve low-frequency structure and motion where possible, and Annealed Noise Correlation (ANC) is introduced to impose temporal coherence on the noise process, suppressing high-frequency jitter without eliminating the flexibility needed for large edits (Kulikov et al., 18 Mar 2026).

4. Similarity Guided Aggregation

SGA modifies how multiple candidate edit directions are combined during early ODE steps. Instead of naively averaging per-sample edit velocities, DynaEdit treats each noise sample as proposing a distinct global edit direction and evaluates that direction by projecting it all the way to the final time (Kulikov et al., 18 Mar 2026). At timestep $f^{\text{tar}}$ 0, with current edit state $f^{\text{tar}}$ 1, DynaEdit forms per-sample projected edits

$f^{\text{tar}}$ 2

Each projected edit is then compared to the source video through a similarity score

$f^{\text{tar}}$ 3

which is converted to a softmax weight with temperature $f^{\text{tar}}$ 4,

$f^{\text{tar}}$ 5

The projected edits are aggregated as

$f^{\text{tar}}$ 6

and the effective update velocity is recovered by

$f^{\text{tar}}$ 7

The stated intuition is that, among candidate target-directed motions, DynaEdit prefers those whose hypothetical final video remains most similar to the source in aspects that should not change, such as camera motion and unaffected objects (Kulikov et al., 18 Mar 2026). The temperature $f^{\text{tar}}$ 8 acts as an alignment-strength parameter: small $f^{\text{tar}}$ 9 makes the aggregation nearly argmax and strongly preserves source trajectory, whereas larger $x^{\text{edit}}$ 0 allows more averaging and more deviation from the source (Kulikov et al., 18 Mar 2026).

In the implementation reported for the method, multi-sample SGA is used only in the first few timesteps, where coarse motion is determined; later steps use a single sample (Kulikov et al., 18 Mar 2026). This design is tied directly to the problem diagnosis: early steps decide low-frequency spatio-temporal structure, whereas later steps largely refine detail. The ablation analysis states that naive averaging remains insufficient because it can blur edits when candidate directions disagree and does not favor the best-aligned global trajectory. Cosine similarity in feature space is reported to outperform MSE, especially for motion alignment (Kulikov et al., 18 Mar 2026).

5. Annealed Noise Correlation and the full DynaEdit pipeline

ANC governs the temporal correlation structure of the noise used to form noisy source states. Rather than drawing independent noise at each timestep, DynaEdit maintains a running noise state $x^{\text{edit}}$ 1 for each sample and updates it through a Markovian mixture of previous noise and fresh Gaussian noise with a timestep-dependent correlation coefficient $x^{\text{edit}}$ 2 (Kulikov et al., 18 Mar 2026). The schedule is monotone increasing: $x^{\text{edit}}$ 3 starts at $x^{\text{edit}}$ 4 at $x^{\text{edit}}$ 5, increases linearly until it reaches $x^{\text{edit}}$ 6 at $x^{\text{edit}}$ 7, and then remains $x^{\text{edit}}$ 8 down to $x^{\text{edit}}$ 9 (Kulikov et al., 18 Mar 2026).

This schedule yields low correlation in early steps and high correlation in later steps. The early regime preserves exploration, which is useful when the model must alter coarse global structure; the late regime stabilizes fine detail and suppresses temporal jitter (Kulikov et al., 18 Mar 2026). The paper contrasts this with two alternatives. A non-Markovian schedule that correlates all steps with a fixed global noise induces ghosting, and a Markovian schedule with decreasing correlation increases jitter rather than reducing it. The reported ablations therefore identify increasing-correlation ANC as the most effective balance between flexibility and temporal stability (Kulikov et al., 18 Mar 2026).

The full DynaEdit loop combines ANC and SGA. The system initializes $V$ 0, sets $V$ 1 at $V$ 2, updates the correlated noise, constructs noisy source and target states, computes source- and target-conditioned velocities, forms per-sample edit directions, aggregates them by SGA, and then applies the ODE update

$V$ 3

This procedure is described as preserving low-frequency alignment through SGA and suppressing high-frequency jitter through ANC, thereby making it possible to edit actions, interactions, and global dynamics without training and without interventions in model internals (Kulikov et al., 18 Mar 2026). A plausible implication is that DynaEdit’s main novelty lies not in replacing the inversion-free flow-editing paradigm, but in controlling its stochastic degrees of freedom so that large dynamic edits remain source-coupled.

6. Editing capabilities, evaluation, and comparative results

DynaEdit is evaluated on four edit categories: object insertion with 2-sided interaction, object swap with different dynamic outcome, action/motion change, and global spatio-temporal effects (Kulikov et al., 18 Mar 2026). Representative examples include inserting an obstacle so a horse jumps instead of running straight, replacing a strawberry with a feather so it falls slowly and floats, inserting a helicopter whose spotlight illuminates the town, adding a flag that one astronaut picks up while another remains unaffected, turning water into lava-like flow, and introducing earthquake-like shaking while preserving camera trajectory (Kulikov et al., 18 Mar 2026). The method is explicitly described as being able to modify only the dynamics of a replaced object while leaving another similar object unchanged, and to make inserted objects affect the scene rather than merely coexist with it (Kulikov et al., 18 Mar 2026).

The benchmark assembled for the method contains 71 triplets of source video, source prompt, target prompt, and optional edited first frame, collected from Pexels at 832×480, 16 fps, and 49–81 frames (Kulikov et al., 18 Mar 2026). Edited first frames for tasks requiring them were produced with Gemini 2.5 Flash Image; otherwise the original first frame was used (Kulikov et al., 18 Mar 2026). Automatic evaluation uses Gemini 3 Pro as a VLM judge with three $V$ 4-to- $V$ 5 scores: content preservation, text adherence, and visual quality. Human evaluation uses 32 participants and approximately 2400 pairwise judgments over 18 representative edits and three main baselines (Kulikov et al., 18 Mar 2026).

Across the 71-edit dataset, the best DynaEdit configuration is reported at approximately 4.21/5 for text adherence, 4.50/5 for content preservation, and up to 3.83/5 for visual quality (Kulikov et al., 18 Mar 2026). The same summary places Runway Gen-4 Aleph at approximately 4.18/5 for text adherence, 4.18/5 for content preservation, and 3.61/5 for visual quality; FlowEdit around 3.85/5, 4.07/5, and 2.87–2.9/5; FlowAlign with weaker content preservation and often very low visual quality; I2V sampling with surprisingly decent text adherence and strong visual quality but weaker source preservation; and ODE inversion lower across all three metrics (Kulikov et al., 18 Mar 2026). The qualitative and category-wise conclusions are that DynaEdit is best overall among training-free methods, comparable to or better than Aleph in the aggregate, dominant over training-free baselines in all categories, better than Aleph on insertion and swap tasks, comparable on action change, and slightly weaker than Aleph on global-effects visual quality while remaining strong overall (Kulikov et al., 18 Mar 2026).

The user study reports that participants preferred DynaEdit over FlowEdit and I2V sampling by a large margin across content preservation, text adherence, and visual quality, and preferred DynaEdit over Aleph more often in content preservation and text adherence while judging it competitive in visual quality (Kulikov et al., 18 Mar 2026). The ablations align with the method’s design rationale: removing SGA worsens motion and camera alignment; removing ANC yields visible jitter; cosine similarity is more effective than MSE for SGA; prompt phrasing is relatively robust; and image-to-video conditioning stabilizes color, identity, and layout compared with text-to-video conditioning without a first frame (Kulikov et al., 18 Mar 2026).

The primary backbone reported for DynaEdit is WAN 2.1 14B I2V, with additional qualitative tests on HunyuanVideo 1.5 I2V, illustrating the method’s model-agnostic character (Kulikov et al., 18 Mar 2026). The default schedule uses $V$ 6, $V$ 7 for the first three timesteps and $V$ 8 thereafter, and the ANC schedule described above (Kulikov et al., 18 Mar 2026). Four hyperparameter configurations are reported: $V$ 9, $dZ_t = V(Z_t, t)\, dt, \quad t \in [0,1].$ 0, $dZ_t = V(Z_t, t)\, dt, \quad t \in [0,1].$ 1, and $dZ_t = V(Z_t, t)\, dt, \quad t \in [0,1].$ 2, corresponding respectively to weaker or stronger classifier-free guidance and looser or stronger alignment to the source (Kulikov et al., 18 Mar 2026).

The method’s limitations are described in direct relation to the base image-to-video model. If the base model has limited resolution, weak rendering of small objects, or poor physical reasoning, DynaEdit inherits those weaknesses. Reported failure modes include distorted small faces, low-quality monsters or other complex objects, and missing or incorrect physical outcomes such as wind effects on trees during storms (Kulikov et al., 18 Mar 2026). Very large spatio-temporal changes can produce either overly conservative edits or unnecessary changes to regions that should remain fixed. The method is also not parameter-free: choosing among the four CFG/ $dZ_t = V(Z_t, t)\, dt, \quad t \in [0,1].$ 3 settings is often useful in practice. In addition, DynaEdit does not incorporate an explicit physics or causal model, and its operating range is bounded by the base model’s resolution and clip-length limits (Kulikov et al., 18 Mar 2026).

The name “DynaEdit” also appears as a broader design frame in adjacent arXiv work. “DynaVis: Dynamically Synthesized UI Widgets for Visualization Editing” describes a system that blends natural language and persistent dynamically synthesized widgets for Vega-Lite editing, and is characterized as almost exactly the kind of system “DynaEdit” aspires to be in visualization editing (Vaithilingam et al., 2024). “DNAEdit: Direct Noise Alignment for Text-Guided Rectified Flow Editing” defines a training-free image-editing pipeline based on Direct Noise Alignment and Mobile Velocity Guidance (Xie et al., 2 Jun 2025). “DynaDrag: Dynamic Drag-Style Image Editing by Motion Prediction” introduces a predict-and-move framework for drag-based image editing with motion prediction, motion supervision, and dynamic handle selection (Sui et al., 2 Jan 2026). “Dynamic Retriever for In-Context Knowledge Editing via Policy Optimization” presents a black-box-compatible dynamic retriever for in-context knowledge editing of LLMs (Nafee et al., 24 Oct 2025). “Dynamic Dyck and Tree Edit Distance: Decompositions and Reductions to String Edit Distance” uses “DynaEdit-style frameworks” to discuss dynamic structured-data editing in algorithmic terms (Das et al., 20 Oct 2025). This suggests that “DynaEdit” functions not only as the title of a specific video-editing method, but also as a broader organizing idea for dynamic, model-mediated editing systems across modalities.