An editing signal execution control method for non-inversion streaming video editing

Abstract

The application provides an editing signal execution control method for non-inversion streaming video editing. It relates to the technical field of video editing. In each reverse sampling step, the source branch and the target branch are constructed according to the source video latent variable, the source prompt, the target prompt and the current sampling time, and a pre-trained speed predictor is called to obtain the original speed difference editing signal between the target branch speed field and the source branch speed field; then the current original editing signal is fused with the displacement editing direction carried by the last sampling step through a causal direction maintaining mechanism to form a step-stable carrying direction; and then the response intensity of the original editing signal is estimated to obtain the execution support field, and the decay control is performed on the weak response area to obtain the actual editing channel input of the sampler. The application can reduce the motion drift, identity flicker and unedited area disturbance in non-inversion video editing, and maintain the target text prompt alignment capability.

Description

Technical Field

[0001] This invention relates to the field of video editing technology, and in particular to a method for controlling the execution of editing signals for inversion-free streaming video editing. Background Technology

[0002] Existing video editing methods can be broadly categorized into inversion-based and non-inversion-based methods. Inversion-based methods typically map the source video back to the noise or latent trajectory of the generative model, and then perform target editing near that trajectory. These methods utilize the trajectory information of the source video, but the computation is computationally intensive, and inversion errors can affect the editing quality. Non-inversion-based methods, on the other hand, attempt to directly generate editing signals during the sampling process using source and target cues, avoiding sample-by-sample optimization or trajectory inversion of the source video. Therefore, they offer better efficiency and ease of use.

[0003] Existing methods primarily focus on constructing, modifying, or guiding the editing signal itself, such as through attention control, feature propagation, optical flow constraints, noise modulation, or velocity difference modulation to improve editing results. While these methods can improve the quality of some video editing, they typically do not treat "signal estimation" and "signal execution" as two independent processes, and lack causal control mechanisms for online velocity difference closed-loop execution. Especially in scenarios without inversion, the model cannot rely on information from future sampling steps, nor can it perform global correction for each step using the source video's specific inversion trajectory. Therefore, a lightweight, causal, and training-independent execution control method is needed to stably use the velocity difference editing signal without altering the pre-trained model.

[0004] Based on the above analysis, this invention proposes an editing signal execution control method for inversion-free streaming video editing. This method preserves the geometric meaning of the original velocity difference editing signal while introducing causal direction preservation and response-aware attenuation during the execution phase. The former is used to fuse the current signal with the historical carried direction, reducing directional abrupt changes between consecutive sampling steps; the latter is used to determine the editing support region based on the current signal response strength, reducing the injection strength in weak response regions. In this way, this invention transforms the original direct execution strategy into an interpretable local optimization execution strategy, providing a stable sampling control mechanism for inversion-free streaming video editing. Summary of the Invention

[0005] The technical problem this invention aims to solve is that in text-guided, inversion-free streaming video editing, although the original velocity difference editing signal has an endpoint displacement interpretation under shared noise geometry, it is an online closed-loop signal during actual neural sampling. If the velocity difference at each step is directly used as the sampler input, it easily amplifies step fluctuations and weak response perturbations, causing motion drift, identity flickering, structural inconsistencies, and distortion of unedited regions. This invention aims to provide an editing signal execution control method that does not change the pre-trained generative model, requires no additional training, and does not require source video inversion trajectories, enabling the velocity difference editing signal to be injected in a more stable, conservative, and response-region-compliant manner during sampling.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: an editing signal execution control method for inversion-free streaming video editing, the specific steps of which are as follows:

[0007] S1. Dual-branch construction steps: Construct a source branch and a target branch based on the source video latent variables, source cues, target cues, and the current sampling time; wherein, the source branch is determined by the source video latent variables, the sampling time, and the noise endpoint, and is used to provide a reference velocity field consistent with the source video; the target branch is obtained by superimposing the current editing latent variable relative to the source video latent variable on the source branch, and is used to provide an online editing velocity field under the target cues condition;

[0008] S2. Original Editing Signal Estimation Step: Call the same pre-trained velocity predictor to calculate the velocity field of the source branch under the source cue condition and the velocity field of the target branch under the target cue condition, and use the difference between the two as the original velocity difference editing signal;

[0009] S3. Causal Direction Preservation Step: Through the causal direction preservation mechanism, the original velocity difference edit signal of the current sampling step is fused with the carrying direction propagated in the previous sampling step to form a step-stable carrying direction; wherein, the update of the carrying direction depends only on the information of the current and past sampling steps, satisfying the causal constraint.

[0010] S4. Response Support Field Estimation Step: Estimate the execution support field based on the response intensity of the original velocity difference edit signal, and perform attenuation control on the weak response region; The execution support field converts the response intensity into an execution coefficient between zero and one, which is used to perform non-amplified attenuation on the edit signal;

[0011] S5. Editing channel input generation step: Multiply the carrying direction by the execution support field point by point to obtain the actual injected sampler editing channel input; wherein, the execution coefficient of the strong response position is close to one, so that the carrying direction can pass through fully; the execution coefficient of the weak response position is close to zero, so that the carrying direction is significantly attenuated;

[0012] S6. Sampling Advancement Step: Using the current latent variables of the edited video and the input of the edited channel, advance to the next sampling step through the sampler transfer operator;

[0013] Repeat steps S1 to S6 until the entire backsampling time series is completed, and the final edited video latent variables are obtained.

[0014] The above steps are performed in the kth sampling step of the backsampling process.

[0015] Furthermore, the causal direction preservation mechanism described in step S3 is implemented in the following way:

[0016] Maintain a carry direction variable that propagates across sampling steps, initially set to zero; starting from the first sampling step, the carry direction is obtained by weighted fusion of the original velocity difference edit signal of the current sampling step and the carry direction of the previous sampling step, and a stable edit direction is retained in the form of exponential memory.

[0017] Furthermore, the estimation of the support field in step S4 includes:

[0018] The absolute value of the original velocity difference editing signal is averaged along the channel dimension to obtain the response intensity of each non-channel potential location. After local smoothing of the response intensity, the Otsu thresholding method is used to adaptively determine the segmentation threshold from the current response distribution. Then, the response intensity is transformed into an execution support field between zero and one using the sigmoid soft assignment function.

[0019] Furthermore, when the response intensity statistics degrade and cannot form a meaningful segment, the execution support field is set to all one, causing the execution cost term to be temporarily disabled.

[0020] Furthermore, the edit channel input of the actual injection sampler is obtained by solving a local convex optimization objective to obtain a closed-form optimal solution; the local convex optimization objective includes: a current signal tracking term, used to respect the editing evidence given by the pre-trained model in the current sampling step; a history direction preservation term, used to constrain the execution input from being completely dominated by single-step fluctuations; and an execution cost term, used to transform the response support structure into a position-dependent injection cost.

[0021] Furthermore, when both memory strength and execution cost are zero, the closed-form optimal solution degenerates into direct execution, that is, the original velocity difference edit signal is used as the complete sampler input; when the execution cost is zero and the memory strength is greater than zero, it degenerates into maintaining only the causal direction; when the memory strength is zero and the execution support field is valid, it degenerates into responding only to support decay.

[0022] Furthermore, the carrying direction variable has the same tensor size as the video latent variable, or it adopts a broadcast form in the channel dimension, spatial dimension, or temporal dimension.

[0023] Furthermore, the estimation method of the execution support field can be replaced by any of the following: quantile threshold, adaptive mean threshold, local contrast threshold, support estimation obtained with the assistance of attention map or saliency map or segmentation mask, or composite support field obtained by multiplying the edit mask provided by the user with the response support field.

[0024] Furthermore, in multi-object or long video editing scenarios, multiple directional variables are maintained for different object regions, different time segments, or different semantic categories, and corresponding local injection inputs are generated according to their respective response support fields. Then, multiple local injection inputs are combined into the final editing channel input.

[0025] Furthermore, the method is applicable to different sampler discretization forms, including Euler steps, improved Euler steps, or higher-order ODE solvers; control is performed once in each main sampling step or substep, or the support field is updated every few sampling steps while the orientation is still updated in each step.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. This invention explicitly distinguishes between editing signal estimation and editing signal execution, improving the stability of inversion-free video editing without changing the pre-trained model and the original speed difference calculation method.

[0028] 2. This invention reduces directional abrupt changes between consecutive sampling steps by maintaining causal orientation, thereby suppressing motion drift and identity flicker caused by fluctuations in online speed difference.

[0029] 3. This invention attenuates weak response locations by responding to the perception execution support field, thereby reducing unexpected disturbances in the background, unedited areas, and areas with insufficient target evidence.

[0030] 4. The execution rules of this invention are derived from a locally convex optimization objective, possessing a closed-form solution, low computational cost, and easy to embed into existing samplers. Fifth, this invention does not require additional training data, fine-tuning processes, or independent model evaluation, making it suitable as a training-independent plugin for existing streaming video editing systems. Attached Figure Description

[0031] Figure 1 This is an example diagram illustrating the video editing effect of the present invention;

[0032] Figure 2 This is a schematic diagram of the editing signal execution control of the present invention;

[0033] Figure 3 This is a flowchart illustrating the editing signal execution control of the present invention;

[0034] Figure 4This is a comparison chart of the qualitative effects of the present invention with FlowDirector, FLATTEN, and other methods in video editing tasks;

[0035] Figure 5 Example 1 is added to the appendix;

[0036] Figure 6 Example 2 is added to the appendix;

[0037] Figure 7 Example 3 is added to the appendix. Detailed Implementation

[0038] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are only used to explain the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Without departing from the spirit and essence of the present invention, those skilled in the art can adjust the relevant variable names, sampler forms, threshold estimation methods, response field smoothing methods, and execution parameter values, and all such adjustments should fall within the scope of protection of the present invention.

[0039] A method for controlling the execution of editing signals in inversion-free streaming video editing, the details of which are as follows:

[0040] 1: Edit the construction of the signal execution controller.

[0041] Figure 1 The invention provides an intuitive demonstration of its effects in a typical text-guided video editing task. An example image illustrating the video editing effect of the invention shows the correspondence between four sets of source videos (dog, car, bird, and cat) and the target edited video. This illustrates that the invention can maintain the layout, camera movement, and continuous frame correspondence of the source videos when performing object replacement, category changes, and appearance modifications.

[0042] Each set of examples includes a source video frame sequence and a target edit frame sequence. The target cues only change local semantic content, while the inter-frame motion, camera angle, and background layout in the source video remain consistent. This phenomenon corresponds to the technical objective of this invention: the edit signal should not be injected as an indiscriminate global perturbation, but should be confined to a location with a clear target response during the execution phase, and maintain a stable direction between consecutive sampling steps.

[0043] This embodiment is based on a pre-trained streaming video generation model. The inputs include source video latent variables, source cues, target cues, backsampled time series, a pre-trained velocity predictor, and a sampler transition operator. The source video latent variables can be the spatiotemporal latent tensor output by the video encoder, or latent variables within the generator used to represent the source video content. Source cues describe the source video content, and target cues describe the desired edited video content. The pre-trained velocity predictor outputs a conditional velocity field given the latent variables, sampling time, and textual conditions.

[0044] In the k-th sampling step, the controller first constructs a source branch and a target branch. The source branch, determined by the source video latent variables, sampling time, and noise endpoints, provides a reference velocity field consistent with the source video. The target branch, obtained by superimposing the source branch with the offset of the current edit latent variables relative to the source video latent variables, provides an online edit velocity field under the target cue conditions. This dual-branch structure allows the controller to compare velocity differences under source and target conditions in the same sampling step without requiring inversion optimization of the source video.

[0045] After obtaining the two branches, the controller calls the same pre-trained velocity predictor to calculate the source cue conditional velocity and the target cue conditional velocity, respectively. The difference between the two is the original velocity difference edit signal. This signal maintains the basic definition of existing streaming editing methods and is still directly provided by the pre-trained model. This invention does not modify the velocity predictor structure, nor does it introduce additional trainable networks, thus it does not destroy the original model's generative priors.

[0046] The original velocity difference edit signal then enters the causal direction preservation unit. This unit maintains a carry direction variable that propagates across sampling steps. Initially, the carry direction is set to zero; starting from the first sampling step, the carry direction is obtained by fusing the current original velocity difference and the carry direction of the previous sampling step. Since the current carry direction depends only on current and past information, this update satisfies the causality requirement and is suitable for embedding in the backsampling process. The role of this unit is not to average all historical signals, but to preserve a stable edit direction in an exponential memory form, so that the execution input does not change drastically with local noise or weak semantic fluctuations in each step.

[0047] The controller calculates the response support field by averaging the absolute values ​​of the original velocity difference edit signal across the channel dimension to obtain the response intensity at each non-channel potential location. This response intensity reflects the magnitude of the edit induced by the target cue relative to the source cue at that location. The controller can perform a light local smoothing of the response intensity to reduce the impact of isolated noise on support estimation; then, it uses the Otsu thresholding method to adaptively determine the segmentation threshold from the current response distribution; finally, it transforms the response intensity into an execution support field between zero and one using a sigmoid soft assignment function. If the response statistics degenerate and cannot form a meaningful segment, the support field can be set to all one, temporarily disabling the execution cost term and thus avoiding erroneous suppression of the effective edit signal.

[0048] After generating the support field, the controller multiplies it point-by-point with the carrying direction to obtain the actual edit channel input injected into the sampler. The support coefficient at strong response locations is close to one, allowing the carrying direction to pass through sufficiently; the support coefficient at weak response locations is close to zero, significantly attenuating the carrying direction. Because the support coefficient is limited to between zero and one, the response support unit does not amplify the edit direction at any location. This design gives the controller a conservative execution characteristic, meaning it only performs strong edits where the target cue provides sufficient evidence of a response, while maintaining the source video content stable where evidence is insufficient.

[0049] The sampler transition operator advances to the next sampling step using the current edited video latent variables and the edit channel input. This process is repeated over the entire backsampled time series until the final edited video latent variables are obtained. The final latent variables can then be transformed into an edited sequence of video frames by the video decoder. Throughout this process, the controller only changes the edit channel input passed to the sampler, without altering the pre-trained velocity predictor, the source and target cues, or introducing any additional training steps.

[0050] Figure 3 The flowchart of the editing signal execution control of this invention illustrates the complete process from input source video latent variables, constructing source and target branches, calling the pre-trained velocity model, calculating the original editing signal, generating the carrying direction, estimating the response support field, forming the injection input, to the propulsion sampler.

[0051] The process begins with the latent variables of the source video and text prompts, constructing the original velocity difference edit signal through source and target branches. This signal is used to update the carry direction and estimate the response support field. The carry direction and response support field are combined to obtain the injection input. The sampler updates the edited latent variables based on this input and proceeds to the next sampling step. This closed-loop process embodies the core idea of ​​this invention: "estimated invariant, execution controlled".

[0052] Figure 2 This is a schematic diagram of the editing signal execution control of the present invention, illustrating the differences between ideal shared noise geometry, online closed-loop execution, and FlowTrack execution control. It is used to explain that although the speed difference has the meaning of endpoint displacement, it still needs to be transformed into a stable execution input through causal direction maintenance and response sensing injection in actual sampling. Figure 2 The geometric and control interpretation behind this process is further given, namely that the velocity difference is only used as an online observation signal, and the final input injected into the sampler is the execution input constrained by both the carrying direction and the support field.

[0053] 2: The relationship between execution goals and special cases.

[0054] This embodiment illustrates that the execution control rules of the present invention can be derived from local execution objectives. The current signal tracking term, the historical orientation preservation term, and the execution cost term in the local objective correspond to three execution requirements. The current signal tracking term ensures that the controller respects the editing evidence given by the pre-trained model in the current sampling step; the historical orientation preservation term ensures that the controller is not completely dominated by single-step fluctuations; and the execution cost term transforms the response support structure into position-dependent injection costs. Since this objective is strictly convex with respect to the input to be executed, a unique closed-form optimal solution exists for each sampling step.

[0055] When both memory strength and execution cost are zero, closed-loop degenerates into direct execution, treating the original velocity difference signal as the complete sampler input. This corresponds to existing uncontrolled execution methods, with the advantage of a direct target semantic response, but the disadvantage of injecting both temporal fluctuations and weak response perturbations. When the execution cost is zero and the memory strength is greater than zero, closed-loop degenerates into maintaining only the causal direction. This smooths out changes in step direction, but still performs direction editing across the entire latent variable range, failing to adequately suppress weak response regions. When the memory strength is zero and the execution support field is valid, closed-loop degenerates into attenuating only the response support. This suppresses perturbations in weak response regions, but lacks sufficient control over cross-sampling step direction fluctuations. The complete FlowTrack incorporates both causal direction maintenance and response support attenuation, thus achieving a balance between directional stability and spatial response selectivity.

[0056] In practical applications, the memory strength parameter can be set according to the needs of the editing task. For relatively localized and continuous appearance editing such as color and material, a smaller or medium memory strength can be used to maintain target cue responsiveness; for object replacement, object removal, or motion-sensitive editing tasks, the memory strength can be appropriately increased to enhance consistency across frames and steps. The soft-assigned sharpness of the response support field is used to control the smoothness of the support boundary. A smaller sharpness produces a smoother support field, suitable for appearance editing with unclear boundaries; a larger sharpness more clearly distinguishes between high-response and low-response areas, suitable for local object editing.

[0057] The execution objective of this invention also has a probabilistic interpretation. The current raw velocity difference can be considered as an observation of the executable input, the direction carried in the previous sampling step as a temporal prior, and the execution cost as a zero-centered conservative injection prior. Minimizing the local execution objective is equivalent to finding the maximum a posteriori execution input under that local Gaussian model. This interpretation shows that this invention is not a simple heuristic filtering, but rather a deterministic, closed-form execution decision obtained by integrating current evidence, historical direction, and response support in each sampling step.

[0058] Performance verification based on the above:

[0059] This embodiment uses the FiVE-Bench video editing evaluation protocol to verify the present invention. FiVE-Bench evaluates video editing quality from multiple perspectives, including source motion preservation capability, target text alignment capability, fidelity of unedited areas, and structural deviation. Among these metrics, higher values ​​for MFS, CLIP, PSNR, and SSIM are better, while lower values ​​for LPIPS, MSE, and Structure are better. Table 1 shows the results under the Overall setting. It can be seen that the present invention achieves superior results in motion preservation, unedited area fidelity, and structural consistency metrics, indicating that execution control can effectively mitigate source content disturbances caused by direct execution.

[0060] Table 1 Comparison of Overall FiVE-Bench Results

[0061]

[0062] To further verify the role of the two execution control components in this invention, a structural ablation experiment was conducted in this embodiment. Direct execution means that causal orientation preservation and response support attenuation are not used; causal orientation only means that only step-carry orientation is preserved; support only means that only response-aware attenuation is preserved; the complete method preserves both. The results are shown in Table 2.

[0063] Table 2 Structural ablation results

[0064]

[0065] As shown in Table 2, maintaining the causal direction alone can improve stability-related indicators more directly, while response support decay alone can significantly reduce errors in the unedited region and structural deviations. The complete method further achieves optimal or near-optimal results on multiple indicators. These results indicate that maintaining the causal direction and response support decay have complementary effects: the former addresses the problem of directional fluctuations during step execution, while the latter addresses the problem of weak spatial response perturbations. When combined, the execution of the edit signal is no longer simply following the current velocity difference, but becomes a stable control input constrained by both historical direction and response evidence.

[0066] Furthermore, the technical effect of this invention is also reflected in the indirect stabilization of editing boundaries. Traditional direct execution methods typically treat the velocity difference field as a uniform and reliable editing command, but different locations in the video latent variables do not have the same editing evidence. The response intensity of the target object, background region, occluded region, and low-texture region in the velocity field may differ significantly. By transforming the response intensity into an execution support field, this invention enables the editing signal to be concentrated at locations with strong target responses and naturally weakened at locations lacking response evidence, thereby reducing problems such as background coloring, replacement of non-target objects, and distortion of motion trajectories. At the same time, causal orientation preservation does not eliminate new evidence in the current sampling step, but rather forms a balance between current evidence and historical orientation, thus avoiding insufficient target semantics caused by excessive smoothing. With the combination of these two aspects, the sampler can both execute the changes required by the target text and maintain the source video structure in both the temporal and spatial dimensions.

[0067] like Figure 4 As shown, Figure 4 This is a qualitative comparison chart of the present invention with methods such as FlowDirector and FLATTEN in video editing tasks, demonstrating the structure preservation and response area control effects of the present invention in changing sports car material, motorcycle color, clothing color, and aircraft category.

[0068] In scenarios such as changes in the material of sports cars, changes in the color of motorcycles, changes in the color of Santa Claus's costume, and replacement of fighter jets with helicopters, this invention, compared to baseline methods with direct or weak control, can better maintain the consistency of the local structure, background content, and unedited areas of the source video while achieving the target semantic changes. Figure 4 The response diagram also shows that the editing response is mainly concentrated in the target object or target attribute area, which is consistent with the design of the present invention to inject the position by supporting field constraints through response.

[0069] The working principle of this invention remains consistent across different editing types. In object replacement tasks, the response support field is typically concentrated near the object being replaced, and maintaining causal orientation helps preserve the object's outline and motion continuity. In color or material editing tasks, the support field can cover the object's appearance area, and carrying orientation ensures that color or material changes remain consistent across consecutive frames. In object addition or deletion tasks, the support field can limit the scope of the addition or removal signal, reducing its impact on the background structure. Therefore, this invention is not a rule specific to a fixed editing category, but rather a general execution control framework for online speed difference editing signals.

[0070] Figure 5 Example 1, which is a supplement to the appendix, shows editing tasks such as turning a swan into a duck, a silver car into a silver jeep, adding sunglasses to a dog, and turning a bicycle purple, to illustrate the applicability of the present invention to the modification of animals, vehicles, and appearance attributes.

[0071] Figure 6 Example 2, which is a supplement to the appendix, shows editing tasks such as adding a hat to a brown bear, changing the appearance and category of a character, turning an eagle into a blue bird, and changing a silver SUV into a pink SUV. This is used to illustrate the generalization effect of the invention in scenes involving characters, animals, and vehicles.

[0072] Figure 7 Example 3, which is a supplement to the appendix, shows editing tasks such as cows turning into horses, boats turning into kayaks, roller skaters turning into Batman, and skiers turning their pants red, to illustrate the ability of this invention to maintain the source motion in category replacement, character replacement, and local changes in clothing.

[0073] Figures 5 to 7 Supplementary qualitative examples from the paper's appendix are provided as supplementary figures. Figure 5 This covers animal category replacement, vehicle category replacement, and color attribute modification. Figure 6 It covers changes in character appearance, animal color, and vehicle color; Figure 7 The coverage includes category replacement, character replacement, and partial changes in clothing. The supplementary examples above correspond to the quantitative results in Tables 1 and 2, demonstrating that the execution control of this invention is not only effective for a single example, but can reduce drift and non-target disturbances in different editing types through the combination of "carrying direction + support attenuation".

[0074] Engineering deployment and alternative implementations:

[0075] In engineering deployments, this invention can be inserted as an execution control module within the sampler into existing inversion-free streaming editing pipelines. For existing systems, only two steps—carrying direction updates and supporting field modulation—need to be added before the original velocity difference edit signal is fed into the sampler. The network structure, weight parameters, text encoder, video decoder, and basic time step scheduling of the pre-trained velocity predictor can remain unchanged. The advantage of this approach is low integration cost and that it does not compromise the prior learning capabilities of the original generative model on large-scale data.

[0076] The carrying direction variable can have the same tensor size as the video latent variable, or it can be broadcast in the channel dimension, spatial dimension, or temporal dimension. If system resources are sufficient, the carrying direction can be maintained on the complete spatiotemporal latent variable to obtain the finest-grained execution control; if the system needs to reduce video memory usage, the support field can be maintained only on the compressed feature, low-resolution latent variable, or keyframe response map, and upsampled to the size required by the sampler before injection. None of the above different implementations change the core mechanism of this invention, namely, the current editing signal must first undergo causal direction preservation, then response support modulation, and only then enter the sampler as an executable input.

[0077] The estimation method for the response support field can also be replaced according to the specific system. Besides the Otsu threshold, quantile thresholds, adaptive mean thresholds, local contrast thresholds, or support estimates obtained with the assistance of attention maps, saliency maps, and segmentation masks can be used. As long as the support estimate can output execution coefficients between zero and one and be used for non-amplified attenuation of the carrying direction, the same technical effect as this invention can be achieved. For application scenarios involving explicit user editing masks, the user mask can also be multiplied by the response support field to further restrict the editing area.

[0078] In multi-object or long video editing scenarios, this invention can be extended to grouped execution control. Specifically, multiple carrying direction variables can be maintained for different object regions, different time segments, or different semantic categories, and corresponding local injection inputs can be generated based on their respective response support fields. Then, the multiple injection inputs are combined into the final editing channel input. This extended form can reduce interference between objects when editing multiple objects simultaneously, and can also maintain the continuity of editing directions in different time segments in long videos. Whether a single carrying direction or multiple carrying directions are used, it is essentially a causal control of the online speed difference editing signal execution stage.

[0079] This invention is also applicable to different sampler discretization methods. Regardless of whether the sampler uses Euler steps, improved Euler steps, higher-order ODE solvers, or other discretization methods, as long as there is an injectable edit channel input in each sampling step, the method of this invention can be used. In higher-order solvers, control can be executed once in each main sampling step, or control can be repeatedly executed in sub-steps; when it is necessary to reduce computational load, the support field can be updated every few sampling steps, while the carrying direction is still updated step by step. The above scheduling methods do not affect the basic scope of protection of this invention.

[0080] From a computational complexity perspective, the additional overhead of this invention mainly comes from point-by-point weighting of the carrying direction, channel averaging of the response intensity, threshold estimation, and point-by-point multiplication. These operations are all lightweight tensor operations, with minimal overhead compared to a single forward inference operation of a pre-trained video velocity model. Since this invention does not introduce backpropagation, perform parameter fine-tuning, or call additional discriminant models, the actual runtime is still primarily determined by the base video generator. Therefore, this invention is particularly suitable for text-guided video editing scenarios requiring rapid experimentation, batch editing, or resource-constrained deployment.

[0081] In terms of safe and compliant use, this invention enhances the controllability of editing execution and the ability to preserve the source video. This capability can be used for legitimate purposes such as film and television post-production, educational content adaptation, accessibility video generation, data augmentation, and low-cost creative editing. At the same time, any high-quality video editing technology can be misused for misleading videos, unauthorized human image modification, or copyright content rewriting. Therefore, in actual product deployment, this invention can be used in conjunction with watermarking, source recording, editing trace annotation, content review, and access control mechanisms to reduce the potential risk of misuse. This description does not change the technical solution of this invention itself, but rather provides a preferred application environment.

[0082] In summary, this invention achieves stable execution of online velocity difference editing signals by incorporating a lightweight execution controller within the inversion-free streaming video editing sampler. This method does not alter the pre-trained model, does not rely on additional training, and does not require source video inversion trajectories, making it a modular approach applicable to existing streaming video editing frameworks. Any technical solutions based on the ideas of this invention that incorporate causal direction preservation, response support estimation, execution cost attenuation, or their equivalent closed-loop execution control before the velocity difference editing signal injection should fall within the scope of protection of this invention.

Claims

1. A method for controlling the execution of editing signals in inversion-free streaming video editing, characterized in that, Includes the following steps: S1. Dual-branch construction steps: Construct a source branch and a target branch based on the source video latent variables, source cues, target cues, and the current sampling time; wherein, the source branch is determined by the source video latent variables, the sampling time, and the noise endpoint, and is used to provide a reference velocity field consistent with the source video; the target branch is obtained by superimposing the current editing latent variable relative to the source video latent variable on the source branch, and is used to provide an online editing velocity field under the target cues condition; S2. Original Editing Signal Estimation Step: Call the same pre-trained velocity predictor to calculate the velocity field of the source branch under the source cue condition and the velocity field of the target branch under the target cue condition, and use the difference between the two as the original velocity difference editing signal; S3. Causal Direction Preservation Step: Through the causal direction preservation mechanism, the original velocity difference edit signal of the current sampling step is fused with the carrying direction propagated in the previous sampling step to form a step-stable carrying direction; wherein, the update of the carrying direction depends only on the information of the current and past sampling steps, satisfying the causal constraint. S4. Response Support Field Estimation Step: Estimate the execution support field based on the response intensity of the original velocity difference edit signal, and perform attenuation control on the weak response region; The execution support field converts the response intensity into an execution coefficient between zero and one, which is used to perform non-amplified attenuation on the edit signal; S5. Editing channel input generation step: Multiply the carrying direction by the execution support field point by point to obtain the actual injected sampler editing channel input; wherein, the execution coefficient of the strong response position is close to one, so that the carrying direction can pass through fully; the execution coefficient of the weak response position is close to zero, so that the carrying direction is significantly attenuated; S6. Sampling Advancement Step: Using the current latent variables of the edited video and the input of the edited channel, advance to the next sampling step through the sampler transfer operator; Repeat steps S1 to S6 until the entire backsampling time series is completed, and the final edited video latent variables are obtained. The steps S1 to S6 described above are executed in the kth sampling step of the backsampling process.

2. The editing signal execution control method for inversion-free streaming video editing according to claim 1, characterized in that, The causal direction preservation mechanism described in step S3 is implemented in the following way: Maintain a carry direction variable that propagates across sampling steps, initially set to zero; starting from the first sampling step, the carry direction is obtained by weighted fusion of the original velocity difference edit signal of the current sampling step and the carry direction of the previous sampling step, and a stable edit direction is retained in the form of exponential memory.

3. The editing signal execution control method for inversion-free streaming video editing according to claim 1, characterized in that, The estimation of the support field in step S4 includes: The absolute value of the original velocity difference editing signal is averaged along the channel dimension to obtain the response intensity of each non-channel potential location. After local smoothing of the response intensity, the Otsu thresholding method is used to adaptively determine the segmentation threshold from the current response distribution. Then, the response intensity is transformed into an execution support field between zero and one using the sigmoid soft assignment function.

4. The editing signal execution control method for inversion-free streaming video editing according to claim 3, characterized in that, When the response intensity statistics degrade and cannot form a meaningful segment, the execution support field is set to all one, causing the execution cost term to be temporarily disabled.

5. The editing signal execution control method for inversion-free streaming video editing according to claim 1, characterized in that, The edit channel input of the actual injection sampler is obtained by solving the local convex optimization objective to obtain the closed-form optimal solution; The local convex optimization objective includes: a current signal tracking term, which respects the editing evidence given by the pre-trained model at the current sampling step; a history orientation preservation term, which constrains the execution input from being completely dominated by single-step fluctuations; and an execution cost term, which transforms the response support structure into a position-dependent injection cost.

6. The editing signal execution control method for inversion-free streaming video editing according to claim 5, characterized in that, When both memory strength and execution cost are zero, the closed-form optimal solution degenerates into direct execution, that is, the original velocity difference edit signal is used as the complete sampler input; when the execution cost is zero and the memory strength is greater than zero, it degenerates into maintaining only the causal direction; when the memory strength is zero and the execution support field is valid, it degenerates into responding only to support decay.

7. The editing signal execution control method for inversion-free streaming video editing according to claim 1, characterized in that, The carrying direction variable has the same tensor size as the video latent variable, or it adopts a broadcast form in the channel dimension, spatial dimension, or temporal dimension.

8. The editing signal execution control method for inversion-free streaming video editing according to claim 1, characterized in that, The estimation method of the execution support field can be replaced by any of the following: quantile threshold, adaptive mean threshold, local contrast threshold, support estimation obtained with the assistance of attention map or saliency map or segmentation mask, or composite support field obtained by multiplying the edit mask provided by the user with the response support field.

9. The editing signal execution control method for inversion-free streaming video editing according to claim 1, characterized in that, In multi-object or long video editing scenarios, multiple directional variables are maintained for different object regions, different time segments, or different semantic categories. Corresponding local injection inputs are generated based on their respective response support fields, and then multiple local injection inputs are combined into the final editing channel input.

10. The editing signal execution control method for inversion-free streaming video editing according to claim 1, characterized in that, The method is applicable to different sampler discretization forms, including Euler steps, improved Euler steps, or higher-order ODE solvers; control is performed once in each main sampling step or substep, or the support field is updated every few sampling steps while the orientation is still updated with each step.

Images