Dual optimization method and system for testing few-step image diffusion model

Abstract

The invention discloses a dual optimization method and system for a few-step image diffusion model during testing, and belongs to the technical field of generative artificial intelligence. On the basis of fixed sampling steps in a few-step image diffusion model reasoning stage, a double-channel optimization mechanism of submerged space alignment and prompt vector alignment is introduced in the reasoning process at the same time: on one hand, an image estimation value generated by a submerged space is utilized in each step of sampling, gradient optimization is performed on submerged space vectors after reward model evaluation, and the image estimation value is obtained; local detail correction is realized; on the other hand, micro-prompt prefixes and suffixes are constructed on the basis of original prompt vectors, the prompt vectors are dynamically optimized and enhanced through reward feedback, and global semantic guidance is achieved. According to the double-alignment mechanism, the quality of the output image, the aesthetic score and the human preference consistency are remarkably improved on the premise of not increasing reasoning steps and not modifying diffusion model parameters. The method has the remarkable advantages that extra training is not needed, compatibility is high, and the method can be directly applied to the reasoning stage.

Description

A test-time dual optimization method and system for few-step image diffusion models Technical Field

[0001] This invention relates to the field of generative artificial intelligence technology, and in particular to a test-time dual optimization method and system for improving the generation quality of a few-step image diffusion model. Background Technology

[0002] Diffusion models, as one of the mainstream model architectures in generative artificial intelligence, have been widely used in text-to-image generation tasks. Traditional multi-step diffusion models typically require twenty or more sampling steps to generate high-quality images through progressive denoising, exhibiting excellent performance in semantic consistency, image detail, and visual aesthetics. However, their high computational cost and long inference time make it difficult to meet real-time and low-latency requirements.

[0003] To reduce generation overhead, researchers have proposed few-step diffusion models, which typically require only one to eight steps to complete generation, significantly improving inference efficiency. However, due to the reduced sampling steps, the generated results still suffer from issues such as missing image details, decreased semantic consistency, lower aesthetic quality compared to multi-step diffusion models, and reduced human preference ratings. Existing methods often rely on model distillation, additional networks, or structural optimization to improve the performance of few-step diffusion models, but these methods often require large amounts of training data and computational resources, making them difficult to apply directly in the inference stage. Furthermore, they place high demands on the model architecture and exhibit poor transferability.

[0004] In recent years, Test-Time Alignment (TTA) has emerged as a novel approach to improve generation quality during the inference phase without additional training. Its core idea is to evaluate the generated results using a pre-trained reward or scoring model and feed the gradients back to differentiable variables (such as latent space vectors) in the generation process, thereby dynamically optimizing the generation trajectory. Existing TTA research almost entirely focuses on multi-step diffusion models. However, multi-step diffusion models are inherently slow inference, and test-time optimization further increases computational overhead. Furthermore, their latent space estimation has limited approximate accuracy for the generated image, providing insufficiently accurate supervision signals and limiting optimization effectiveness. In contrast, short-step diffusion models are not only efficient inference but also provide more accurate estimations of the final image, offering more reliable supervision signals for test-time optimization. Moreover, the generation performance of diffusion models is influenced not only by the latent space trajectory but also by the text cue vector, which plays a crucial role in the semantic consistency, detail representation, and aesthetic quality of the image. While existing research has explored optimizing cue vectors at test time, most methods primarily target multi-step diffusion models or optimize the latent space or cue vectors individually, making it difficult to fully leverage the synergistic effect of both in short-step generation scenarios.

[0005] Based on the above analysis, there is an urgent need for a dual-path test-time optimization technique that can simultaneously optimize the latent space and cue vectors during the inference phase. This would fully leverage the advantages of few-step diffusion models in terms of generation efficiency and latent space prediction accuracy, achieving high-quality, low-latency image generation. This technique can be directly deployed without additional training and significantly improves visual quality and semantic consistency through joint optimization of the latent space and cue vectors, providing a feasible solution for the practical application of few-step diffusion models. Summary of the Invention

[0006] To address the issues of limited generation quality, decreased semantic consistency, and insufficient visual aesthetics in existing few-step diffusion models during the inference phase, this invention provides a test-time dual optimization method and system for few-step image diffusion models. By jointly optimizing the latent space vector and cue vector during the inference phase and combining it with a reward model for dynamic guidance, high-quality, low-latency image generation is achieved, and text-image consistency is improved.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, the present invention proposes a dual optimization method for testing a few-step image diffusion model, comprising: receiving a text prompt input by a user, wherein the text prompt is a natural language description of the image to be generated; generating an original prompt vector based on the text prompt, and constructing a differentiable prompt prefix vector and a prompt suffix vector, concatenating the three to form an enhanced prompt vector, and simultaneously initializing the latent space vector of the few-step diffusion model; executing the few-step diffusion model generation according to a preset sampling step, wherein the number of few-step sampling steps is 1-8; in each sampling step, outputting an image estimate based on the current latent space vector and the enhanced prompt vector; using a reward model to perform quality evaluation on the image estimate, obtaining a comprehensive reward signal, and performing backpropagation optimization on the current latent space vector, and the prompt prefix vector and prompt suffix vector in the enhanced prompt vector based on the comprehensive reward signal; the optimized latent space vector and enhanced prompt vector are used in the next sampling step; after completing all sampling steps, generating an output image based on the finally optimized latent space vector.

[0008] Furthermore, the reward model includes at least two combinations of an aesthetic rating predictor, a CLIP relevance prediction model, and an image preference rating model, for generating images based on a comprehensive evaluation of aesthetic quality, text relevance, and human preference.

[0009] Furthermore, the prompt prefix vector and prompt suffix vector are randomly initialized and have the following shape: ,in For the number of additional tokens, For text embedding dimensions.

[0010] Furthermore, the output image estimate includes: inputting the current latent space vector and the enhanced cue vector into the denoising network of the diffusion model to calculate the estimated final latent space vector; and decoding the estimated final latent space vector through a variational autoencoder to obtain the image estimate in the pixel space.

[0011] Furthermore, by limiting the magnitude of each gradient update or applying regularization constraints, perturbation magnitude constraints are imposed on the updates of the cue prefix and cue suffix vectors, thus constraining the update vectors. The norm does not exceed the preset threshold.

[0012] Furthermore, the optimization of the latent space vector and the enhanced cue vector includes: calculating the gradient of the latent space vector based on the comprehensive reward signal, and performing gradient ascent update on the latent space vector; calculating the gradient of the enhanced cue vector based on the comprehensive reward signal, and performing gradient ascent update on the cue prefix vector and the cue suffix vector, while keeping the original cue vector part fixed.

[0013] Furthermore, the comprehensive reward signal is calculated by combining mean reward, weighted reward, or exponential decay reward strategies, and the weight of the weighted reward is preset according to the importance of aesthetic score, text relevance, and human preference score.

[0014] Furthermore, latent space optimization and cue vector optimization are performed synchronously in each sampling step.

[0015] Secondly, this invention proposes a test-time dual optimization system for a few-step image diffusion model, which implements the aforementioned test-time dual optimization method for a few-step image diffusion model.

[0016] The beneficial effects of this invention are as follows: By simultaneously optimizing the latent space and cue vector during the inference phase of a few-step diffusion model, and combining this with the guidance and synchronous update mechanism of a reward model, high-quality, low-latency image generation is achieved. Compared with existing technologies, this invention can significantly improve image detail, text consistency, and visual aesthetic quality while maintaining generation efficiency. Furthermore, it can be directly deployed without additional training, demonstrating good practicality, versatility, and robustness. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 is an overall flowchart of a dual optimization method for improving the generation quality of a few-step image diffusion model according to the present invention during testing; Figure 2 is a flowchart of the processing of the dual optimization mechanism (synchronous update of latent space and cue vector) in the present invention; Figure 3 is a schematic diagram of the module structure of the system of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of the invention.

[0020] This invention proposes a test-time dual optimization method for a few-step image diffusion model. This method does not require any modification to the pre-trained weights of the diffusion model. Instead, it dynamically and synchronously corrects latent space noise and enhances cue vectors by utilizing the instantaneous feedback provided by the reward model during the inference phase.

[0021] First, the original process of the few-step diffusion model for generating images is introduced: (1) First, a text prompt is input, such as "a cat on the grass". The text is converted into an original prompt vector by a pre-trained text encoder. An initial latent space vector is generated by sampling from the standard normal distribution, which serves as the starting point for image generation. The latent space vector is the noise data that represents the compressed representation of the image in the diffusion model.

[0022] (2) Set a fixed number of sampling steps T, and start from the initial time step (t=T) to perform the denoising operation step by step. At each step, use a denoising network of the diffusion model (such as U-Net structure) to predict noise and update the latent space state based on the current latent space vector and cue vector. Repeat this process, reduce the noise ratio at each step, and gradually restore the image structure.

[0023] (3) After completing all sampling steps (t=0), the final latent space vector is decoded and converted into an image output in pixel space. The original process relies on the model's pre-trained parameters and has no dynamic optimization mechanism. As shown in Figure 1, the dual optimization method for testing the few-step image diffusion model proposed in this invention mainly includes steps such as input text prompts and initialization, cyclic sampling and evaluation, dual parameter optimization, and final decoding and generation. This process is particularly suitable for few-step generation models (such as SD3.5-Turbo, SDXL-Turbo, etc.), aiming to significantly improve the aesthetic quality and semantic consistency of the generated image while maintaining low latency characteristics.

[0024] (1) System initialization and enhanced prompt construction Referring to the starting part of Figure 1, the present invention first receives the text prompts input by the user and the preset number of diffusion model sampling steps. , The process typically involves 1 to 8 steps; and includes initialization of differentiable prefix and suffix cue vectors and latent space vectors: on the latent space side, according to the requirements of the diffusion model, starting from the standard normal distribution... The initial latent space vector is obtained by sampling. As the starting point of the image generation process, this vector is the object of operation in the subsequent latent space optimization module.

[0025] On the text side, the original text prompt is input into a pre-trained diffusion model text encoder module (such as CLIP or T5 encoder) to obtain the original prompt embedding vector. To achieve test-time optimizability of the cue vector, this invention further constructs learnable continuous embedding vectors as an additional feature to enhance the cue: i.e., having shape Hint prefix vector and shape Hint suffix vector ,in Embed the spatial vector dimension of the text. , These are the number of additional tokens added to the cue prefix vector and cue suffix vector, respectively, both with a value of 8. These two additional vectors are typically initialized with random values ​​or text embeddings representing general aesthetic guidelines. Finally, the enhanced cue vector... It consists of three parts: The original prompt is embedded within. It remains fixed during the reasoning process, while and These are the key parameters that can be updated by gradients in subsequent inference steps.

[0026] (2) During testing, image prediction and reward calculation are performed in a few-step sampling loop. In each step or selected optimization step, this step is responsible for quickly obtaining image quality feedback based on the current parameters.

[0027] At the current time step As shown in the judgment logic in Figure 1, the system determines whether the current sampling step count is less than the maximum set step count (i.e., whether all steps have been completed). If not, it enters the loop to perform image estimation and evaluation. The dual optimization processing mechanism shown in Figure 2 adjusts the current latent space vector... and the current enhanced cue vector Denoising network input to diffusion model The time step can be derived from the formulas for forward noise addition and reverse denoising in the diffusion model. The final latent variable vector at time The calculation formula for the prediction is as follows. The calculation formula here varies depending on the underlying mathematical mechanism of the diffusion model: For models based on Score Matching: For flow matching-based models: .

[0028] Subsequently, the potential space was estimated. The image is decoded into a pixel-space image using a pre-trained variational autoencoder (VAE). The estimated final pixel-space image, along with the original text prompt, is then input into a predefined reward model. The system calculates the comprehensive reward signal. Reward Model The system is a multi-module integrated system, including an aesthetic scoring model (Aesthetic model) to evaluate image aesthetics, a CLIP similarity scoring model to evaluate image-text semantic consistency, and preference models (HpsV2 model, PickScore model, ImageReward model) to evaluate alignment with human preferences. It also integrates reward signals. The weighted average is calculated by combining the mean, weighted sum, and / or exponential decay of each sub-reward item, and is used to balance the importance of each reward. The weights are preset based on the importance of aesthetic score, text relevance, and human preference score. In this embodiment, the weight for text relevance is set to 0.4, and the weights for aesthetic score and human preference score are set to 0.3.

[0029] (3) Dual-parameter optimization based on reward gradient: This step constitutes the core mechanism of this invention, aiming to transform the high-level semantic evaluation signal into the low-level parameter correction instruction through the closed-loop feedback structure shown in Figure 2. In the current sampling step... In the process, the system first calculates the comprehensive reward signal based on the aforementioned steps. This invention constructs an objective function to guide the optimization. To transform the reward maximization problem into a standard gradient descent optimization problem, this invention defines an optimization loss function. .

[0030] Subsequently, using the automatic differentiation mechanism of the PyTorch deep learning framework, the loss function is accurately calculated through backpropagation using the chain rule. The gradient is relative to the gradients of the two key optimizable parameter sets: the latent space gradient. and the gradient of the hint vector Based on the calculated gradient field, the system executes a synchronous dual optimization update strategy: for latent space vector optimization, gradient ascent update is performed. .in The preset latent space step size. This update process directly affects the latent representation of the image, and its physical meaning is to fine-tune the noise distribution along the direction of reward increase. Since the latent space preserves the spatial structure information of the image, this operation focuses on local refinement of the currently generated image, such as correcting object edges, filling in missing textures, or eliminating unnatural artifacts, thereby significantly improving the image fidelity and detail.

[0031] For optimizing the augmentation cue vector: perform gradient ascent update. .in The step size for the cue vector is specified here. Note that only the prefix and suffix parts of the enhanced cue vector are updated; the original cue portion in the middle remains frozen. This operation operates on the embedding space of the text encoder, focusing on providing global semantic guidance for the generation process. By adjusting the embedding vectors of the cue prefix and suffix, the system can dynamically guide the generation trajectory towards a distribution region with higher aesthetic scores or better human preferences.

[0032] In addition to gradient ascent updates, those skilled in the art can also use momentum optimization, Adam, or other differentiable optimization methods for updates, which will not be illustrated here.

[0033] Furthermore, to prevent "semantic drift" (i.e., the generated embedding vector deviates from the effective representation space of natural language) in continuous optimization, this embodiment introduces a perturbation constraint mechanism. Specifically, the system calculates the update amount after each update. and its application Norm truncation or dynamic magnitude clipping ensures that the updated vector Always located centered on the original initialization vector - Within the neighborhood, thus ensuring the safety of optimization and the stability of semantics.

[0034] (4) Fewer-step denoising sampling and final generation: After completing the dual optimization of the above parameters, the system re-injects the parameter state carrying higher quality information into the generation process, returns to the main loop shown in Figure 1, and executes the operation to advance to the next sampling time step. This stage realizes a smooth switch from the "optimization domain" to the "generation domain".

[0035] Specifically, the optimized latent space vector and optimized enhanced cue vectors The input is fed into the diffusion model, where a pre-defined denoising sampler performs a single-step operation to obtain the latent space vector for the next time step. The system then returns to the decision step, repeating the image estimation, evaluation, and optimization process described above until the sampling loop is complete. When the termination condition is met (i.e., sampling ends), as shown in the output section of Figure 1, the system decodes the final latent space vector to generate the final high-quality image.

[0036] To verify the effectiveness of the proposed dual-path optimization method during testing in a few-step diffusion model, 500 text prompts from the HPSv2 dataset were selected as the test set, and systematic experiments were conducted on the SD3.5 series models based on the Flow Matching framework and the SDXL series models based on the Score Matching framework. In the experiments, SD3.5 and SDXL served as the basic multi-step diffusion models, with a fixed sampling step count of 28 steps; SD3.5-Turbo and SDXL-Turbo served as the few-step inference baselines, with a fixed sampling step count of 4 steps. All models maintained their original parameters, with the proposed dual-path synchronous optimization mechanism for latent space and prompt vectors introduced only during the inference stage to ensure that the experimental results accurately reflect the actual effect of the proposed method in the testing-time optimization scenario. In a unified experimental environment, a corresponding image was generated for each prompt, and the generated results were quantitatively evaluated using five mainstream indicators: Aesthetic score, CLIP image-text matching degree, HPSv2 preference score, ImageReward preference score, and PickScore preference score, in order to comprehensively measure the aesthetic performance, semantic consistency, and consistency with human preferences of the image.

[0037] The experimental results on the SD3.5 series models are shown in Table 1. Compared to the 28-step SD3.5 model, the 4-step SD3.5-Turbo shows a slight decrease in aesthetic quality and semantic consistency. However, when the dual optimization mechanism of this invention is introduced into SD3.5-Turbo, its Aesthetic, CLIP, HPSv2, ImageReward, and PickScore are all significantly improved. The overall performance is not only significantly better than the basic Turbo model, but also surpasses the multi-step sampling SD3.5 in most metrics. For example, HPSv2 improved from 0.283 to 0.316, ImageReward from 1.198 to 1.468, and PickScore from 0.209 to 0.289, demonstrating the ability of this invention to achieve or even surpass the level of multi-step diffusion models while maintaining extremely low latency in few-step inference.

[0038] Table 1. Comparison of the generation quality of this invention on the SD3.5 series models. The experimental results on the SDXL series models are shown in Table 2. The 4-step SDXL-Turbo exhibits a certain degree of performance degradation compared to the 28-step SDXL. However, after introducing this invention, all five metrics of SDXL-Turbo are comprehensively improved. This not only recovers the performance lost during the acceleration process of the Turbo model but also significantly outperforms the basic SDXL in key dimensions such as semantic relevance and preference scoring. Specifically, CLIP relevance increased from 0.269 to 0.291, HPSv2 from 0.277 to 0.309, ImageReward from 1.086 to 1.460, and PickScore from 0.203 to 0.282, with stable and significant improvements, further validating the universality and robustness of the method in different diffusion model systems.

[0039] Table 2 Comparison of the generation quality of the present invention on the SDXL series models The experiments demonstrated above show that this invention can significantly improve the generation quality of the few-step diffusion model without altering the original model structure or introducing additional training costs, relying solely on immediate reward feedback during the inference phase. Latent space optimization enhances the structural consistency of local details, while cue vector optimization further strengthens image-text semantic alignment. The synchronous update mechanism of both enables the Turbo model to achieve stable and significant comprehensive improvements across multiple dimensions, including aesthetic quality, semantic relevance, and human preference ratings. Extensive experiments show that this invention exhibits excellent model independence, applicable not only to existing mainstream models such as SDXL and SD3.5, but also demonstrating good transferability and practical application value to other diffusion frameworks based on few-step inference.

[0040] This embodiment also provides a test-time dual optimization system for a few-step image diffusion model, which implements the aforementioned method. As shown in Figure 3, this system adopts a modular design and mainly works collaboratively during inference. It includes: a sampling execution module, which serves as the main control center of the system, responsible for initializing and maintaining the few-step sampling loop, and scheduling the optimization process and the execution of the denoising solver; a latent space resolution module, which receives the current latent space vector and the enhanced cue vector, predicts the latent space of the denoised image, and decodes it to generate a pixel estimation image; a reward evaluation module, which receives the image estimate and the original input text, integrates a multi-dimensional evaluation model, calculates the comprehensive reward signal, and generates feedback information to guide parameter correction; a latent space vector optimization module, which receives the feedback information, performs parameter correction based on the reward signal, achieves efficient fine-tuning of image details and local structures, and feeds back the optimized latent space vector to the sampling execution module; and a cue vector optimization module, which receives the feedback information, synchronously performs parameter correction based on the reward signal, achieves global semantic guidance and aesthetic style optimization, and has a built-in perturbation constraint mechanism to maintain semantic stability, and feeds back the optimized enhanced cue vector to the sampling execution module.

[0041] In this embodiment, the latent space parsing module generates an initial image estimate by decoding the latent space vector, and further adjusts the generation process by combining conditional noise or control signals to improve the diversity and quality of the generated images.

[0042] For the system embodiments, since they basically correspond to the method embodiments, relevant details can be found in the descriptions of the method embodiments; the implementation methods of the remaining modules will not be repeated here. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0043] The system embodiments of the present invention can be applied to any device with data processing capabilities, such as a computer or other similar device. The system embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution.

[0044] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. Those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A test-time dual optimization method for a few-step image diffusion model, characterized in that, include: Receive text prompts input by the user, wherein the text prompts are natural language descriptions of the image to be generated; The original cue vector is generated based on the text prompts, and a differentiable cue prefix vector and cue suffix vector are constructed. These three vectors are concatenated to form an enhanced cue vector. Simultaneously, the latent space vector of the few-step diffusion model is initialized. The few-step diffusion model is generated according to a preset sampling step, with the number of few sampling steps being 1-8. In each sampling step, an image estimate is output based on the current latent space vector and the enhanced cue vector. The image estimate is quality-assessed using a reward model to obtain a comprehensive reward signal. Backpropagation optimization is then performed on the current latent space vector and the cue prefix and cue suffix vectors in the enhanced cue vector based on the comprehensive reward signal. The optimized latent space vector and enhanced cue vector are used in the next sampling step. After completing all sampling steps, an output image is generated based on the final optimized latent space vector.

2. The dual optimization method for testing a few-step image diffusion model according to claim 1, characterized in that, The reward model includes at least two combinations of an aesthetic rating predictor, a CLIP relevance prediction model, and an image preference rating model, used to generate images based on a comprehensive evaluation of aesthetic quality, text relevance, and human preference.

3. The dual optimization method for testing a few-step image diffusion model according to claim 1, characterized in that, The prompt prefix vector and prompt suffix vector are randomly initialized and have the following shape: ,in For the number of additional tokens, For text embedding dimensions.

4. The dual optimization method for testing a few-step image diffusion model according to claim 1, characterized in that, The output image estimation includes: inputting the current latent space vector and the enhanced cue vector into the denoising network of the diffusion model to calculate the estimated final latent space vector; and decoding the estimated final latent space vector through a variational autoencoder to obtain the image estimation value in the pixel space.

5. The dual optimization method for testing a few-step image diffusion model according to claim 1, characterized in that, By limiting the magnitude of each gradient update or applying regularization constraints, perturbation magnitude constraints are imposed on the updates of the cue prefix and cue suffix vectors, thus constraining the update vectors. The norm does not exceed the preset threshold.

6. The dual optimization method for testing a few-step image diffusion model according to claim 1, characterized in that, The optimization of the latent space vector and the enhanced cue vector includes: calculating the gradient of the latent space vector based on the comprehensive reward signal and performing gradient ascent update on the latent space vector; calculating the gradient of the enhanced cue vector based on the comprehensive reward signal and performing gradient ascent update on the cue prefix vector and the cue suffix vector, while keeping the original cue vector part fixed.

7. The dual optimization method for testing a few-step image diffusion model according to claim 1, characterized in that, The comprehensive reward signal is calculated by combining mean reward, weighted reward, or exponential decay reward strategies. The weight of the weighted reward is preset based on the importance of aesthetic score, text relevance, and human preference score.

8. The dual optimization method for testing a few-step image diffusion model according to claim 1, characterized in that, Latent space optimization and cue vector optimization are performed synchronously in each sampling step.

9. A test-time dual optimization system for a few-step image diffusion model, used to implement the test-time dual optimization method for a few-step image diffusion model as described in claim 1, characterized in that, The system includes: The latent space parsing module is used to generate image estimates based on the latent space vectors. The reward evaluation module is used to generate a comprehensive reward signal for the image estimates. The latent space optimization module is used to update the latent space vectors based on the comprehensive reward signal. The cue vector optimization module is used to construct and update the cue prefix vector and cue suffix vector. The sampling execution module is used to perform few-step diffusion generation according to preset steps and trigger the synchronous update of the latent space optimization module and the cue vector optimization module after each sampling.

10. The dual-optimization system for testing a few-step image diffusion model according to claim 9, characterized in that, The latent space resolution module generates initial image estimates by decoding latent space vectors and adjusts the generation process by incorporating conditional noise or control signals, thereby improving the diversity and quality of the generated images.

Images