A natural language driven video tone mapping parameter and three-dimensional lookup table generation system

Abstract

The application discloses a natural language driven video color adjustment parameter and three-dimensional lookup table generation system, and relates to the technical field of digital image processing.The system comprises an anchor point library interface, a deterministic text embedding, a rule priori, a kernel weighted retrieval, a residual learning and a fusion unit, generates a color adjustment parameter vector through multi-branch fusion, and outputs a structured explanation triple containing a rule contribution, retrieval information and a residual to realize decision traceability.The system also comprises a lookup table export and consistency checking unit, which can generate a.cube format three-dimensional lookup table and verify the consistency of the three-dimensional lookup table with a pixel-by-pixel coloring result.The application does not need to rely on an external network or a pre-trained model, supports offline deterministic reasoning, effectively solves the problems of non-smooth output, non-reproducibility and lack of explainability in traditional schemes, and is suitable for video color adjustment creation and teaching scenes.

Description

Technical Field

[0001] This invention relates to the field of data processing and digital image and video signal processing technology, and in particular to a natural language-driven video color grading parameter and three-dimensional lookup table generation system. Background Technology

[0002] In multimedia art creation and teaching scenarios, the working vocabulary of video color grading is usually expressed in natural language descriptions, while the actual operable parameters on the color grading software side are a limited set of bounded knob parameters (such as color temperature, contrast, saturation, gamma, shadow boost, highlight compression, and overall intensity). These parameter values ​​ultimately need to be baked into a 3D lookup table (3DLUT) and applied to the video footage. Converting creative intentions in natural language form into precise parameter vectors and corresponding 3D lookup tables is a highly repetitive and experience-dependent routine task.

[0003] Existing solutions mainly fall into two categories: one is the preset library retrieval solution, which embeds the input phrase and retrieves the nearest existing preset and directly uses its parameters. Although this solution is low in implementation cost and easy to interpret, the hard nearest neighbor retrieval returns an existing point by construction, causing any phrase with intermediate semantics to be absorbed by a certain discrete preset. When the phrase transitions continuously from one style to another, the output presents a step-like "platform plus jump" rather than a continuous change, which contradicts the smooth interpolation behavior required for parameter control. The other category is the solution that relies on large cloud models or commercial color matching kits. Its deployment cost is high, requiring network dependence, pay-per-use, and unstable latency. Moreover, the output is not guaranteed to be consistent when the same phrase is input twice, which is difficult to meet the requirements in scenarios where teaching demonstrations must be repeatable and audit traceability must be traceable. In addition, there is also an end-to-end regression network solution, which directly regresses parameter vectors from text using a single network. However, its output is a single vector, which does not expose the decision basis and cannot trace the source of the value of a certain dimension branch by branch. Furthermore, when the anchor point sample size is limited, it is necessary to fit a complete absolute working point on all parameter dimensions.

[0004] At the same time, existing solutions typically only output parameter values. Even if they claim to drive lookup tables, they do not provide a means to verify the consistency between the exported lookup table and the local pixel-by-pixel coloring path, which may result in unverified differences between the exported object and the local preview.

[0005] Therefore, there is an urgent need for a system that can achieve offline deterministic mapping from natural language to color parameters and 3D lookup tables without relying on external networks and pre-trained model weights, while also possessing output smoothness, decision traceability, and verifiability of the derived products. Summary of the Invention

[0006] The purpose of this invention is to provide a natural language-driven video color grading parameter and 3D lookup table generation system, which solves the technical problems in the prior art, such as output adsorption and staircase effect caused by hard nearest neighbor retrieval, deployment uncertainty and non-reproducibility caused by cloud service reliance, lack of decision traceability in end-to-end regression networks, and lack of consistency verification means between the exported lookup table and local pixel-by-pixel colorization. These problems make it impossible to achieve smooth, deterministic, interpretable and verifiable mapping of natural language to color grading parameters and 3D lookup tables in an offline environment.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0008] A natural language-driven video color grading parameter and 3D lookup table generation system includes an anchor point library interface unit for loading an anchor point library in a replaceable manner. The anchor point library consists of multiple anchor points, each anchor point containing an anchor point text and a corresponding bounded normalized parameter vector, wherein each component of the parameter vector takes values ​​in [0,1].

[0009] A deterministic text embedding unit is used to generate feature vectors for the input natural language query phrase and all anchor texts in the anchor library using a deterministic character n-tuple hashing method. The unit does not load external pre-trained model weight files and does not initiate external network interface calls.

[0010] The rule prior unit is used to map the vocabulary consisting of direction words and intensity modifiers to a signed offset relative to a neutral default vector. The dimension of the neutral default vector is the same as the dimension of the parameter vector, and each component has a value of 0.5. The offset is output only when an item in the vocabulary is identified in the query phrase, and an explicit abandonment flag is output when no item is identified.

[0011] The kernel-weighted retrieval unit is used to calculate the cosine similarity between the feature vector of the query phrase and the feature vector of each anchor text, and to perform a weighted average of the parameter vectors corresponding to the top k anchors in the similarity ranking using the normalized exponential weight of the temperature parameter T.

[0012] The residual learning unit includes a multilayer perceptron fitted on the anchor point library. The input of the multilayer perceptron is the feature vector of the query phrase. The residual label is the anchor parameter vector minus the sum of the rule prior output and the kernel weighted retrieval output. The training samples consist of anchor text and its corresponding parameter vector, and are used to output the residual part that is not explained by the rule prior unit and the kernel weighted retrieval unit.

[0013] The fusion unit is used to perform a weighted summation of the outputs of the rule prior unit, the kernel weighted retrieval unit and the residual learning unit according to a preset weight. When the rule prior unit outputs a discard flag, it is numerated into an all-zero vector to participate in the summation, and the summation result is component-limited to [0,1] to obtain the fusion parameter vector.

[0014] A configuration selection unit is used to pre-select k and T from the candidate values ​​of the kernel-weighted retrieval unit. The selection is based solely on the average Shannon entropy of the normalized weights obtained by the kernel-weighted retrieval unit in the interpolation query, with the average Shannon entropy not being lower than ln2 as a qualification condition. The selection is not based on any evaluation results of the fusion parameter vector. The interpolation query is generated by concatenating two anchor texts according to a preset ratio to generate query text, or by linearly interpolating the feature vectors of two anchor texts according to the same ratio to generate query vector.

[0015] An explanation output unit is used to output explanation triples on the same path as the output of the fusion parameter vector. The first element of the explanation triple is the triggered rule term and its parameter contribution or the abandonment flag. The second element is the identifier of the selected anchor point, similarity and normalized weight or explicit retrieval closing flag. The third element is the residual part. The explanation triple satisfies that the parameter vector reconstructed by it is consistent with the output fusion parameter vector.

[0016] The lookup table export unit is used to sample the shading driven by the fusion parameter vector at 33^3 grid points of a unit RGB cube and write out a 3D lookup table file in Adobe .cube format;

[0017] The consistency verification unit is used to compare the result obtained by directly processing each pixel according to the coloring with the result obtained by trilinear interpolation through the three-dimensional lookup table file, and output the absolute channel error between the two.

[0018] As an improvement, the feature vector generated by the deterministic text embedding unit is consistent bit-by-bit across different processes; the external network interface call count of the system on the inference path is zero;

[0019] The fusion parameter vector obtained by repeatedly executing the same query phrase at least 100 times is identical in every element, and the fusion parameter vector obtained after the process is restarted is identical in every element.

[0020] The system's inference path dependencies do not include 3D lookup table sampling objects and color difference calculation objects.

[0021] As an improvement, the working configuration selected by the configuration selection unit is that k takes all anchor points in the anchor point library and T takes 2.0;

[0022] The system outputs an evaluation conclusion under the conditions of the working configuration, equal-count ternary band layering, and the deterministic text embedding, and records the working configuration as a condition for the application of the output conclusion.

[0023] As an improvement, a smoothness evaluation unit is also included, for:

[0024] The interpolated query is synthesized between anchor pairs and divided into three bands (near, middle, and far) based on semantic distance. The following four metrics are then calculated on the middle band:

[0025] The adsorption index is equal to one minus the ratio of the L_1 distance from the predicted parameter vector to the nearest anchor parameter vector to the average L_1 distance between any two anchor points.

[0026] Occupancy rate, which is equal to the proportion of predictions falling into the anchor sphere with a radius equal to the median nearest neighbor distance of the anchor cloud;

[0027] Diversity collapse is equal to one minus the proportion of distinct prediction units; the distinct prediction units are judged by rounding each component of the prediction parameter vector to two decimal places, and two prediction vectors are considered distinct if at least one component is different.

[0028] The maximum jump in the path is equal to the maximum single-step change on the multi-step interpolation path between two anchor points; the single-step change is the L_1 distance between two adjacent predicted parameter vectors.

[0029] The smoothness evaluation unit uses hard nearest neighbor retrieval as a pure retrieval reference. The hard nearest neighbor retrieval is a retrieval with k set to 1 and no weighting. The confidence interval of the difference of the four measures is given by the resampling percentile bootstrap method, and the sign consistency rate is given by multiple random seeds.

[0030] As an improvement, the smoothness evaluation unit excludes candidate metrics in the form of nearest neighbor dominance ratio. The basis for the exclusion is that the candidate metric measures the dominance relationship between neighbors rather than predicting whether the collapse will occur to the anchor point.

[0031] Furthermore, the conclusions output by the smoothness evaluation unit are limited to the midband and are not extrapolated to the nearband and the farband.

[0032] As an improvement, the explanation output unit includes an export adaptation layer, which is set outside the inference core and is used to structurally export the identifiers, similarities and normalized weights of each selected anchor point even when the output of the kernel weighted retrieval unit is packaged or replaced by temperature configuration.

[0033] The explanation output unit is also used to output an explicit mark for closed branches instead of omitting the branch;

[0034] Furthermore, the mean absolute error and maximum absolute error between the parameter vector reconstructed from the explained triplet and the output fused parameter vector are both zero.

[0035] As an improvement, an anchor point library admission unit is also included, for:

[0036] Before the anchor point library participates in the kernel-weighted retrieval unit, a decodability test is performed on it. The decodability test includes: for each anchor point in the anchor point library, only its nearest neighbor anchor point is taken in the text embedding space, the L_1 distance between the parameter vector of the anchor point and the parameter vector of the nearest neighbor anchor point is measured, and the in-library statistic of the distance is compared with the permutation null distribution obtained by randomly shuffling the pairing of anchor point text and parameter vector at least 200 times. At the same time, the ridge regression leave-one-out error from the anchor point text feature vector to the anchor point parameter vector is output in parallel. The regularization coefficient of the ridge regression is set to 1.0 and its comparison result with the same permutation null distribution.

[0037] The decodeability test is independent of the output of the fusion unit and is not contingent on the correctness of the anchor parameter vector.

[0038] As an improvement, when the decodability test determines that the anchor point library is undecodable, the anchor point library admission unit records that the text of the anchor point library does not carry decodable parameter information and restricts the use of the anchor point library for the kernel weighted retrieval unit.

[0039] The undecodeable condition includes a first case and a second case. The first case is a corpus in which the anchor text is composed of arbitrary names and the average of the pairwise embedding distances and the type-to-form ratio are not low. The type-to-form ratio is the ratio of the number of distinct terms in the corpus to the total number of terms. "Not low" means that the ratio is not less than 0.8. The second case is a construction in which no corresponding relationship is established between the anchor text and the anchor parameter vector.

[0040] As an improvement, in the anchor point library loaded by the anchor point library interface unit, the parameter vector of at least one anchor point library is constructed in the following manner: the parameter vector is extracted by rejecting sampling, such that the L_1 distance between any two parameter vectors is not less than the preset minimum interval, and then the extracted parameter vector is subjected to axis-by-axis monotonic deformation and first axis permutation.

[0041] Furthermore, the vocabulary of the anchor text in the anchor text library does not intersect with the vocabulary of the rule prior unit, making the hit count of the rule prior unit on the anchor text in the anchor text library zero. The anchor text is generated by a deterministic mapping, in which each axis of the parameter vector contributes a polar term and one of the multi-level intensity terms taken from an independent vocabulary. The anchor text is generated across multiple sentence structures.

[0042] As an improvement, the system freezes resource rejection lines before measurement. These resource rejection lines include a 95th percentile upper limit of 200 milliseconds for single query latency, a peak resident memory upper limit of 521 MB, a fitting time upper limit of 30 seconds, a bitwise consistency requirement, and a zero inference path network interface call count.

[0043] The system's single query latency under the ordinary central processing unit and the deterministic text embedding path does not exceed the 95th percentile limit, the peak resident memory does not exceed the peak resident memory limit, and the fitting time does not exceed the fitting time limit.

[0044] The beneficial effects of this invention are as follows: This system adopts a deterministic text embedding scheme based on character n-tuple hashing, which does not rely on external pre-trained model weights and network calls. It can complete the generation of color adjustment parameters in an offline CPU environment. In most scenarios, it can ensure that the output results of multiple inferences for the same natural language query remain consistent, which is convenient for use in teaching demonstrations and audit traceability scenarios.

[0045] By combining kernel-weighted retrieval with residual learning and rule-based prior multi-branch fusion, compared with traditional hard nearest neighbor retrieval, the problems of output adsorption and step transition can be alleviated to a certain extent, allowing natural language input with continuously changing semantics to obtain relatively smooth transition of color parameter output.

[0046] The system is equipped with an interpretation output unit that can output an interpretation triplet containing rule contribution, anchor point retrieval information, and residual part. This enables structured tracing of the generation source of color grading parameters, facilitating debugging and verification of the contribution of each part to the final color grading result.

[0047] The system supports exporting 3D lookup tables in Adobe .cube format. It also includes a consistency check unit that compares the channel errors between direct pixel-by-pixel coloring and 3D-LUT interpolation, helping to identify potential deviations between the exported lookup table and the original coloring effect. In addition, the anchor point library admission verification mechanism helps to identify whether there is a readable correspondence between the text in the anchor point library and the color adjustment parameters, reducing the search risk caused by invalid anchor point libraries. Attached Figure Description

[0048] Figure 1 This is an overview diagram of the offline inference pipeline of a natural language-driven video color grading parameter and 3D lookup table generation system of the present invention.

[0049] Figure 2 This is a schematic diagram of two smoothness metrics used in the baseline evaluation of this invention.

[0050] Figure 3 This is a comparison and evaluation result diagram of the fully fused scheme of the present invention and the hard 1-NN retrieval.

[0051] Figure 4 This is a schematic diagram showing the comparison results of four metrics in the smoothness evaluation unit of this invention: adsorption index, occupancy rate, diversity collapse, and maximum path jump.

[0052] Figure 5 This is a schematic diagram of the process for exporting and verifying the consistency of the three-dimensional lookup table in this invention. Detailed Implementation

[0053] To make the content of this invention easier to understand, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings.

[0054] Example 1: System Overall Pipeline Implementation

[0055] This embodiment provides a complete implementation flow for offline inference of the system. First, a pre-set anchor library file (JSON format) is loaded through the anchor library interface unit. Each anchor contains text and its corresponding normalized parameter vector. After the system starts, the deterministic text embedding unit generates a fixed-dimensional feature vector (using character 3-gram hashing, FNV-1a hash function, dimension set to 256) for the input natural language query phrase and all anchor text. The rule prior unit parses the direction and intensity terms in the query in parallel. If a match is found, the offset is output; otherwise, a discard flag is output. The kernel weighted retrieval unit calculates the cosine similarity and takes the temperature-normalized weighted average of all anchors (k=N, T=2.0). The residual learning unit (a pre-trained three-layer MLP, 128 neurons per layer) outputs the residuals. The fusion unit sums the residuals by weighting them with α=0.1, β=0.7, and γ=0.2 and limits the amplitude to [0,1] to obtain the final parameter vector. When the rule prior unit outputs an abandonment flag, its value is converted into an all-zero vector and used for weighted summation. The interpretation output unit simultaneously records the rule triggering status, retrieval weights, and residual values, forming an interpretation triplet. Finally, the lookup table export unit samples the color transformation on 33³ grid points to generate a .cube file; the consistency verification unit loads this file and compares it with the pixel-by-pixel processing results, outputting the maximum absolute error for each channel. In this embodiment, the single query latency on a regular CPU is less than 200ms, and there are no network calls throughout the process.

[0056] Example 2: A combined implementation of deterministic text embedding and rule priors

[0057] This embodiment focuses on the text processing stage. The deterministic text embedding unit uses a character 4-gram hash with a fixed feature dimension of 512. It uses the CityHash algorithm with a fixed seed to ensure that the same text generates identical feature vectors in different processes or after a restart. The embedding process does not rely on any external dictionary or pre-trained model. The rule prior unit independently constructs a direction-intensity mapping table. For example, "warm" corresponds to a positive offset of +0.15 on the color temperature axis, "cold" corresponds to -0.15, "slight" has an intensity coefficient of 0.3, and "strong" has a coefficient of 0.8. When querying "slightly warm tone", the rule prior outputs an offset of +0.15 × 0.3 = +0.045 (applying only to the color temperature axis); when querying terms not included in the vocabulary, it outputs None. This embodiment computes the two units in parallel. During a query, feature extraction is performed only once on the text, and anchor text features are cached in memory to avoid redundant computation. In actual testing, the total time for feature extraction and rule matching is less than 5ms.

[0058] Example 3: Implementation of Configuration Selection and Smoothness Evaluation for Kernel-Weighted Retrieval

[0059] This embodiment describes the selection method of k and T in kernel-weighted retrieval and the smoothness verification. Before deployment, a set of interpolation queries is constructed (synthesized between any two anchor texts at ratios of 0.1, 0.2, ..., 0.9). Specifically, the synthesis method is as follows: two anchor texts are concatenated at a given ratio, such as combining 70% of the vocabulary of the first text with 30% of the vocabulary of the second text when the ratio is 0.3, or linearly interpolating the feature vectors of the two anchor texts at the same ratio to generate query vectors. Candidate k values ​​(from 1 to all anchor texts) and T values ​​(from 0.5 to 5.0) are traversed, and the average Shannon entropy of the normalized weights of the kernel-weighted retrieval output under each group is calculated. The smallest T value that satisfies the average entropy ≥ ln2 is selected, and k is taken from all anchor texts. The selected result (e.g., k = all, T = 2.0) is fixed for inference. Subsequently, the smoothness evaluation unit divides the interpolation path between anchor pairs into three bands—near, middle, and far—based on semantic distance. In the middle band, it calculates the adsorption index, occupancy rate, diversity collapse, and maximum path jump, and compares this with hard nearest neighbor retrieval (k=1). A bootstrap method (1000 resampling times) is used to obtain the 95% confidence intervals of the differences. If the sign consistency rate of the four differences is positive, the system's smoothness is considered superior to hard retrieval. This embodiment ensures that the configuration selection does not depend on the fusion output, but is based solely on the retrieval weight entropy value, thus possessing objectivity.

[0060] Example 4: Implementation of Structured Derivation and Reconstruction Consistency of Explained Output

[0061] This embodiment achieves traceable output interpretation. The interpretation output unit sets up an export adaptation layer outside the inference core. This adaptation layer is independent of the temperature parameter adjustment of the kernel-weighted retrieval. Regardless of whether temperature scaling or wrapping is enabled in the kernel-weighted retrieval, the adaptation layer can extract the unique identifier, original similarity value, and normalized weight of the selected anchor point to form the second element. The first element obtains the triggered terms and their vector contributions from the rule prior unit; if it is a rejection, it outputs "-". The third element is directly taken from the output floating-point vector of the residual learning unit. The interpretation triplet reconstructs the parameter vector as follows: Reconstructed value = α × rule contribution (zero vector if any, when rejecting) + β × weighted anchor vector + γ × residual vector. After amplitude limiting, it is completely identical to the parameter vector output by the fusion unit component by component (mean absolute error and maximum absolute error are both 0). This embodiment ensures the reconstruction equation holds by design, facilitating auditing and debugging. The export adaptation layer supports storing the interpretation triplet in JSON format on disk for subsequent analysis.

[0062] Example 5: Implementation of Anchor Point Library Construction and Decodeability Admission Verification

[0063] This embodiment involves the construction and validity verification of the anchor point library. During construction, the rejection sampling method is used to generate anchor point parameter vectors in the [0,1]^d space, requiring that the L1 distance between any two vectors is ≥0.15, and applying sigmoid deformation (deformation coefficients are randomly selected from 0.8 to 1.2) and one random axis permutation to each axis. The anchor point text is mapped to polar words (such as "warm / cold") and intensity words (such as "slight / medium / strong") by each axis of the parameter vector, and combined into various sentence structures (such as "adjust the color tone to a warmer tone" and "increase the degree of warmth"), and ensuring that the text vocabulary has no intersection with the rule prior vocabulary. Before the anchor point library is used for retrieval, a decodability test is performed: for each anchor point, the nearest neighbor anchor point is found in the text embedding space, and the L1 distance between their parameter vectors is measured to obtain the in-library statistics; simultaneously, the text-parameter pairings are randomly shuffled 200 times to generate a zero distribution. If the in-library statistics are significantly lower than the 5th quartile of the zero distribution, and the ridge regression leave-one-out error (with a regularization coefficient of 1.0) is lower than the mean of the zero distribution, then it is determined to be decodable and allowed to be used; otherwise, it is marked as undecodable and the library is restricted to use only for rule priors or residuals, and not for kernel-weighted retrieval. This embodiment ensures that the anchor point library has semantic-parameter correspondences, improving retrieval reliability.

[0064] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A natural language-driven video color grading parameter and 3D lookup table generation system, characterized in that, include: An anchor point library interface unit is used to load an anchor point library in a replaceable manner. The anchor point library consists of multiple anchor points. Each anchor point contains an anchor point text and a corresponding bounded normalized parameter vector. The components of the parameter vector take values ​​in [0,1]. A deterministic text embedding unit is used to generate feature vectors for the input natural language query phrase and all anchor texts in the anchor library using a deterministic character n-tuple hashing method. The unit does not load external pre-trained model weight files and does not initiate external network interface calls. The rule prior unit is used to map the vocabulary consisting of direction words and intensity modifiers to a signed offset relative to a neutral default vector. The dimension of the neutral default vector is the same as the dimension of the parameter vector, and each component has a value of 0.

5. The offset is output only when a word in the vocabulary is identified in the query phrase, and an explicit abandonment flag is output when no word is identified. The kernel-weighted retrieval unit is used to calculate the cosine similarity between the feature vector of the query phrase and the feature vector of each anchor text, and to perform a weighted average of the parameter vectors corresponding to the top k anchors in the similarity ranking using the normalized exponential weight of the temperature parameter T. The residual learning unit includes a multilayer perceptron fitted on the anchor point library. The input of the multilayer perceptron is the feature vector of the query phrase. The residual label is the anchor parameter vector minus the sum of the rule prior output and the kernel weighted retrieval output. The training samples consist of anchor text and its corresponding parameter vector, and are used to output the residual part that is not explained by the rule prior unit and the kernel weighted retrieval unit. The fusion unit is used to perform a weighted summation of the outputs of the rule prior unit, the kernel weighted retrieval unit and the residual learning unit according to a preset weight. When the rule prior unit outputs a discard flag, it is numerically converted into an all-zero vector to participate in the summation, and the summation result is component-limited to [0,1] to obtain the fusion parameter vector. A configuration selection unit is used to pre-select k and T from the candidate values ​​of the kernel-weighted retrieval unit. The selection is based solely on the average Shannon entropy of the normalized weights obtained by the kernel-weighted retrieval unit in the interpolation query, with the average Shannon entropy not being lower than ln2 as a qualification condition. The selection is not based on any evaluation results of the fusion parameter vector. The interpolation query is generated by concatenating two anchor texts according to a preset ratio to generate query text, or by linearly interpolating the feature vectors of two anchor texts according to the same ratio to generate query vector. An explanation output unit is used to output explanation triples on the same path as the output of the fusion parameter vector. The first element of the explanation triple is the triggered rule term and its parameter contribution or the abandonment flag. The second element is the identifier of the selected anchor point, similarity and normalized weight or explicit retrieval closing flag. The third element is the residual part. The explanation triple satisfies that the parameter vector reconstructed by it is consistent with the output fusion parameter vector. The lookup table export unit is used to sample the shading driven by the fusion parameter vector at 33^3 grid points of a unit RGB cube and write out a 3D lookup table file in Adobe .cube format; The consistency verification unit is used to compare the result obtained by directly processing each pixel according to the coloring with the result obtained by trilinear interpolation through the three-dimensional lookup table file, and output the absolute channel error between the two.

2. The natural language-driven video color grading parameter and 3D lookup table generation system according to claim 1, characterized in that, The feature vector generated by the deterministic text embedding unit is consistent bit-by-bit across different processes; the external network interface call count of the system on the inference path is zero; The fusion parameter vector obtained by repeatedly executing the same query phrase at least 100 times is identical in every element, and the fusion parameter vector obtained after the process is restarted is identical in every element. The system's inference path dependencies do not include 3D lookup table sampling objects and color difference calculation objects.

3. The natural language-driven video color grading parameter and 3D lookup table generation system according to claim 2, characterized in that, The configuration selected by the configuration selection unit is that k takes all anchor points in the anchor point library and T takes 2.0; The system outputs an evaluation conclusion under the conditions of the working configuration, equal-count ternary band layering, and the deterministic text embedding, and records the working configuration as a condition for the application of the output conclusion.

4. The natural language-driven video color grading parameter and 3D lookup table generation system according to claim 1, characterized in that, It also includes a smoothness evaluation unit for: The interpolated query is synthesized between anchor pairs and divided into three bands (near, middle, and far) based on semantic distance. The following four metrics are then calculated on the middle band: The adsorption index is equal to one minus the ratio of the L_1 distance from the predicted parameter vector to the nearest anchor parameter vector to the average L_1 distance between any two anchor points. Occupancy rate, which is equal to the proportion of predictions falling into the anchor sphere with a radius equal to the median nearest neighbor distance of the anchor cloud; Diversity collapse is equal to one minus the proportion of distinct prediction units; the distinct prediction units are judged by rounding each component of the prediction parameter vector to two decimal places, and two prediction vectors are considered distinct if at least one component is different. The maximum jump in the path is equal to the maximum single-step change on the multi-step interpolation path between two anchor points; the single-step change is the L_1 distance between two adjacent predicted parameter vectors. The smoothness evaluation unit uses hard nearest neighbor retrieval as a pure retrieval reference. The hard nearest neighbor retrieval is a retrieval with k set to 1 and no weighting. The confidence interval of the difference of the four measures is given by the resampling percentile bootstrap method, and the sign consistency rate is given by multiple random seeds.

5. The natural language-driven video color grading parameter and 3D lookup table generation system according to claim 4, characterized in that, The smoothness evaluation unit excludes candidate metrics in the form of nearest neighbor dominance ratio. The basis for the exclusion is that the candidate metric measures the dominance relationship between neighbors rather than predicting whether the collapse will occur to the anchor point. Furthermore, the conclusions output by the smoothness evaluation unit are limited to the midband and are not extrapolated to the nearband and the farband.

6. The natural language-driven video color grading parameter and 3D lookup table generation system according to claim 1, characterized in that, The explanation output unit includes an export adaptation layer, which is set outside the inference core. It is used to export the identifier, similarity and normalized weight of each selected anchor point in a structured manner, even when the output of the kernel weighted retrieval unit is packaged or replaced by temperature configuration. The explanation output unit is also used to output an explicit mark for closed branches instead of omitting the branch; Furthermore, the mean absolute error and maximum absolute error between the parameter vector reconstructed from the explained triplet and the output fused parameter vector are both zero.

7. The natural language-driven video color grading parameter and 3D lookup table generation system according to claim 1, characterized in that, It also includes an anchor library admission unit, used for: Before the anchor point library participates in the kernel-weighted retrieval unit, a decodability test is performed on it. The decodability test includes: for each anchor point in the anchor point library, only its nearest neighbor anchor point is taken in the text embedding space, the L_1 distance between the parameter vector of the anchor point and the parameter vector of the nearest neighbor anchor point is measured, and the in-library statistic of the distance is compared with the permutation null distribution obtained by randomly shuffling the pairing of anchor point text and parameter vector at least 200 times. At the same time, the ridge regression leave-one-out error from the anchor point text feature vector to the anchor point parameter vector is output in parallel. The regularization coefficient of the ridge regression is set to 1.0 and its comparison result with the same permutation null distribution. The decodeability test is independent of the output of the fusion unit and is not contingent on the correctness of the anchor parameter vector.

8. The natural language-driven video color grading parameter and 3D lookup table generation system according to claim 7, characterized in that, When the decodability test determines that the anchor point library is undecodable, the anchor point library admission unit records that the text in the anchor point library does not carry decodable parameter information and restricts the use of the anchor point library for the kernel weighted retrieval unit. The undecodeable condition includes a first case and a second case. The first case is a corpus in which the anchor text is composed of arbitrary names and the average of the pairwise embedding distances and the type-to-form ratio are not low. The type-to-form ratio is the ratio of the number of distinct terms in the corpus to the total number of terms. "Not low" means that the ratio is not less than 0.

8. The second case is a construction in which no corresponding relationship is established between the anchor text and the anchor parameter vector.

9. The natural language-driven video color grading parameter and 3D lookup table generation system according to claim 8, characterized in that, In the anchor point library loaded by the anchor point library interface unit, the parameter vector of at least one anchor point library is constructed in the following manner: the parameter vector is extracted by rejecting sampling, such that the L_1 distance between any two parameter vectors is not less than the preset minimum interval, and then the extracted parameter vector is subjected to axis-by-axis monotonic deformation and first axis permutation. Furthermore, the vocabulary of the anchor text in the anchor text library does not intersect with the vocabulary of the rule prior unit, making the hit count of the rule prior unit on the anchor text in the anchor text library zero. The anchor text is generated by a deterministic mapping, in which each axis of the parameter vector contributes a polar term and one of the multi-level intensity terms taken from an independent vocabulary. The anchor text is generated across multiple sentence structures.

10. The natural language-driven video color grading parameter and 3D lookup table generation system according to claim 1, characterized in that, The system freezes resource rejection lines before measurement. These resource rejection lines include a 95th percentile upper limit of 200 milliseconds for single query latency, a peak resident memory upper limit of 521MB, a fitting time upper limit of 30 seconds, a bitwise consistency requirement, and a zero inference path network interface call count. The system's single query latency under the ordinary central processing unit and the deterministic text embedding path does not exceed the 95th percentile limit, the peak resident memory does not exceed the peak resident memory limit, and the fitting time does not exceed the fitting time limit.

Images