Diffusion time step adaptive derivation method based on potential space PSNR

Abstract

The invention discloses a diffusion time step self-adaptive derivation method based on a potential space PSNR (Potential Signal to Noise Ratio), which comprises the following steps of: under an extremely low bit rate condition, taking compression potential as a condition, adaptively determining a starting time step of diffusion back-stepping by utilizing the potential domain PSNR, and introducing noise to count matching loss in a training stage, so as to obtain the diffusion time step of the diffusion back-stepping. And empirical statistics of sample differences are aligned with theoretical diffusion noise statistics, so that the perception quality and stability of reconstruction are improved while the code rate bpp is ensured. A diffusion time step matched with the PSNR is deduced in a self-adaptive mode according to the PSNR of potential space calculation, noise characteristic matching loss is constructed through theoretical noise statistics corresponding to the time step, joint optimization is carried out on the noise characteristic matching loss and main loss, and therefore the statistical characteristics of compression / condition noise and diffusion noise are aligned in the training process; reconstruction degradation caused by mismatch is relieved, and the reconstruction quality and stability under the extremely low bit rate are improved.

Description

Technical Field

[0001] This invention belongs to the fields of machine learning, image coding and generative model training technology, and in particular relates to an adaptive derivation method for diffusion time step based on latent space PSNR and its noise statistical characteristic matching constraint in diffusion model training, which is used to improve the reconstruction quality and robustness of conditional generation in compression decoding scenarios. Background Technology

[0002] In the new era of the Internet of Things, image communication technology has deeply penetrated all aspects of social life and industrial development. Whether it's smart homes and intelligent security, industrial automation monitoring, remote medical imaging, or satellite remote sensing, the scale of image data generation and circulation continues to climb. With the explosive growth in the number of connected devices and the exponential expansion of data volume, existing communication links and storage systems face significant pressure: limited bandwidth resources and constrained end-side and cloud-side storage have become key bottlenecks restricting the further development of image communication. Achieving efficient, low-latency, and scalable image transmission and storage under limited resource conditions has become a core problem that the industry urgently needs to solve. Since the digitization of television signals, image compression technology has been a fundamental means of alleviating bandwidth and storage pressures. From early transform coding (such as the JPEG series) to recent deep learning compression (LIC) frameworks, end-to-end learning autoencoders and probabilistic entropy models have significantly improved distortion-rate performance at medium bitrates. However, with the widespread adoption of 5G / 6G and massive IoT terminals, ultra-low bitrate scenarios are becoming increasingly common: cameras need to transmit critical images in real time over narrowband or intermittent links, and industrial and vehicle-to-everything (V2X) systems need to robustly transmit status images under limited bandwidth. In this scenario, traditional methods are prone to significant perceptual degradation and artifacts (blurred details, missing textures, block artifacts, ringing, etc.) under high compression ratios, and distortion metrics such as PSNR and MS-SSIM are insufficient to fully characterize "realism / naturalness." Therefore, perceptual metrics such as LPIPS and FID are gradually being used to evaluate the quality of reconstructed perception and have driven the development of "generative compression."

[0003] Generative models, especially the recently emerging diffusion models, offer a new technical approach for ultra-low bitrate image compression. Diffusion models learn the data distribution through a forward noise addition / reverse denoising process, enabling them to fill in high-frequency details that traditional encoding and decoding methods struggle to preserve during the reconstruction phase, achieving reconstruction quality that better aligns with human perception. In generative compression paradigms, diffusion models are often used as conditional generative decoders: conditioned on compression potential, they begin with an appropriate noise level and perform reverse diffusion, thereby recovering realistic textures and structures at extremely low bitrates. This mechanism not only has potential for improving perceptual quality but also makes it possible to further reduce the bitrate within the same bandwidth budget.

[0004] However, applying diffusion models to extremely low bitrate compression still faces key challenges: First, the choice of the starting time step is crucial to reconstruction quality and computational cost. In real-world systems, the "conditions" often originate from latent or degenerate observations, and their error statistics are not naturally consistent with the Gaussian noise assumed in the diffusion process. Second, when inference primarily occurs in the latent space, a unified and synergistic approach to robustly measuring latent distortion and determining the appropriate diffusion time step remains elusive. Third, even with a obtained time step, the mismatch between conditional noise statistics and diffusion theory noise can still affect the reversibility and stability of the inverse process, requiring explicit constraints or alignment during training.

[0005] In summary, for generative decoding scenarios involving extremely low bitrate image compression, there is an urgent need for a training and inference method that adaptively derives the diffusion time step based on a latent spatial distortion metric and matches the conditional noise statistics with the diffusion theory noise. This method aims to balance reconstruction realism, detail fidelity, and system stability within limited bitrate and computational budgets. Effectively addressing these issues will directly enhance the practical value of diffusion models in IoT and embedded constrained scenarios, promoting the large-scale application of extremely low bitrate image communication. Summary of the Invention

[0006] The present invention aims to adaptively derive the matching diffusion time step based on the PSNR calculated from the potential space, and simultaneously construct a noise characteristic matching loss using the theoretical noise statistics (zero mean and variance determined by the cumulative coefficient) corresponding to the time step, and jointly optimize it with the main loss, thereby aligning the statistical characteristics of compression / conditional noise and diffusion noise during the training process, alleviating the reconstruction degradation caused by mismatch and improving the reconstruction quality and stability at extremely low bit rates.

[0007] This invention proposes a generative decoding method and system based on latent PSNR-adaptive diffusion time step and noise statistical matching. Under extremely low bit rate conditions, this scheme uses compressed latent data as a condition and adaptively determines the starting time step of diffusion backpropagation using the latent domain PSNR. Furthermore, it introduces noise statistical matching loss during the training phase to align the empirical statistics of sample differences with the theoretical diffusion noise statistics, thereby improving the perceived quality and stability of the reconstructed data while maintaining the bit rate (bpp).

[0008] Overall method flow:

[0009] S1 Conditional Acquisition and Bitrate Statistics

[0010] S1.1 Input Generation

[0011] To effectively capture the overall structural information of the image, this method first maps the input image to the VAE (Variational Auto-Encoder) latent space built into the diffusion model. This process is accomplished by the encoder. Extract global semantic features from the image while preserving key structural details such as edges and textures. Its mathematical form is as follows:

[0012] For the original image, For the diffusion model, an auto-encoder encoder, This represents the encoded raw global features.

[0013] S1.2 Feature Compression Module:

[0014] To further improve the efficiency of latent space coding, this method introduces the concept of Information Bottleneck (IB) in the feature compression module to perform structured compression of global features while maximizing the amount of information relevant to the reconstruction task. This module consists of a backbone latent encoder, a quantizer, and a decoder.

[0015] First of all Input with parameters encoder to obtain potential representations :

[0016]

[0017] After quantification Obtaining the latent representation :

[0018]

[0019] Finally, input the parameter. decoder get , representing the global features extracted from the decompression.

[0020]

[0021] To capture the spatial statistical relationships between latent features, this method further introduces a hyperprior structure. The parameter is A hyper-prior encoder. It is used to obtain edge information. To capture Spatial dependencies between elements:

[0022]

[0023] z is determined by the quantizer Quantification :

[0024]

[0025] Will enter with learning parameters decoder Generate entropy parameters .

[0026]

[0027] The advanced prior mechanism enhances the probabilistic modeling capability of latent representations, making compression more efficient and reconstruction more accurate.

[0028] S2 achieves time-step adaptation from the forward-diffusion-calibrated PSNR (PSNR → )

[0029] S2.1 PSNR–t calibration for forward diffusion

[0030] The forward propagation process of the diffusion model gradually injects Gaussian noise into the image. This method utilizes this gradual noise addition process to construct a monotonic correspondence between PSNR and time steps, thereby achieving "time step adaptation" in the inference stage. Given an original clean image... With diffusion schedule

[0031] Its forward (noise-adding) process is as follows:

[0032]

[0033] A set of noise images is generated step by step from 0 to T (where T can be set as needed, and in this patent, T=1000). .in Represents the given raw data , Indicates at time step Noisy data The distribution follows a mean of The standard deviation is The Gaussian distribution.

[0034] For each time step Calculate the MSE and PSNR of the latent domain compared to the original graph:

[0035]

[0036]

[0037] Where N is the total number of pixels. This is the image dynamic range constant. From this, a monotonically non-increasing calibration curve is obtained. :

[0038]

[0039] It can be stored as a lookup table and the corresponding time step can be found using the nearest neighbor interpolation method.

[0040] S2.2 During runtime, the time step is calculated by back-calculating the PSNR to transform the monotonic function of the above equation. Inverse equation:

[0041]

[0042] Based on this, the starting time step is obtained by estimating the image domain PSNR:

[0043]

[0044] In terms of implementation, The basic solution is achieved by combining a lookup table (LUT) with an inverse lookup, while satisfying the boundary conditions: when Time to take ;when Time to take .

[0045] S3 Noise Statistical Matching and Joint Optimization

[0046] Let the training objective be latent. This implementation uses the first four channels:

[0047]

[0048] in Represents the latent space feature tensor. This indicates that the tensor is a four-dimensional real tensor with a size of . B represents the number of input samples in a batch, C represents the number of feature channels, H represents the height of the feature map (pixels / grid size), and W represents the width of the feature map.

[0049] Calculate the mean squared error and PSNR over the latent domain:

[0050]

[0051]

[0052] The time step is obtained by performing a nearest neighbor lookup using the PSNR-t lookup table calibrated offline in S2.

[0053]

[0054] Subsequently The theoretical noise standard deviation was obtained. and variance .

[0055]

[0056]

[0057] Define nonnormalized latent residuals :

[0058]

[0059] Aggregate the tensors in spatial and channel dimensions, and then obtain scalar empirical statistics for all dimensions (b,c,h,w) in a batch:

[0060]

[0061] Joint optimization using noise statistical matching loss , From mean loss With variance loss Adding them together, we get:

[0062]

[0063]

[0064]

[0065] Image Reconstruction with S4 Diffusion Prior

[0066] To achieve high-perceptual-quality reconstruction results at extremely low bitrates, this invention employs a conditional diffusion decoding module. Without compromising existing generative capabilities, content variables are lightweightly injected into the diffusion denoising process, enabling the reconstructed image to possess both global structural consistency and realistic details.

[0067] S4.1 Overview of Stable Diffusion

[0068] First, using the first-stage encoder Image Encoding as a latent representation ,

[0069]

[0070] Then, a Markov chain is used to gradually increase the noise, with the noise intensity adjusted from the default noise parameters. Control, record

[0071]

[0072] This represents the data at diffusion step t. This represents pure noise sampled from a standard normal distribution. This is the cumulative noise retention term.

[0073] Noise estimator with U-Net structure learned during training To predict the added noise, the objective function is:

[0074]

[0075] This represents the loss function of the diffusion model. Represent the expected value of the following random variables: Clean images Text conditions: This patent is empty. For diffusion time step, It is random noise. For the model's output This represents the mean square error of the expected value.

[0076] S4.2 Conditional Diffusion Decoding Module

[0077] To fully utilize the generation prior of pre-trained stable diffusion under extremely low bitrate conditions, this invention performs lightweight conditional injection on the diffusion backbone.

[0078] S4.2.1 Overall Approach and Trainable Range

[0079] While maintaining the stability of the diffusion backbone, a small control branch is introduced to continuously inject conditional features related to the content variables into the diffusion model denoising network.

[0080] S4.2.2 Input and Channel Organization

[0081] The input to the control branch is a noise state. Conditional potential With text conditions It participates in the computation across layers by combining attention and temporal embeddings, and the results are returned to the sampler for use. (Corresponding:)

[0082]

[0083] S4.2.3 Structure and Injection Location

[0084] The Control Module is isomorphic to the Stable Diffusion backbone U-Net. At multiple layers, 1×1 zero-convolutions (Zero-Conv) are used to project the control module features onto the corresponding channels of the backbone and then inject them using residual summation. Three lists of zero-convolutions are maintained to achieve stable multi-point conditional fusion: layer-by-layer injection on the encoding side, single-point injection at bottlenecks, and layer-by-layer backfeeding on the decoding side.

[0085] S5 Reverse Diffusion Decoding Process

[0086] During the inference phase, an interval sampler is used to sample data at predetermined steps (steps = 50). By uniformly selecting locations from high noise to low noise and performing back-reasoning, the potential reconstruction can be obtained. Finally, the VAE decoder is used to... Reconstructed image :

[0087] .

[0088] Joint Objective of S6 Training Phase

[0089] During the training phase, the following joint loss is used for end-to-end optimization:

[0090]

[0091] in:

[0092] The reconstruction loss is calculated as "L_1 / L_2" in the pixel domain and / or latent domain.

[0093] Perceived loss (which may include LPIPS, DISTS, etc.);

[0094] For rate distortion;

[0095] This is a noise statistics matching term. Attached Figure Description

[0096] Figure 1 It is an image compression module.

[0097] Figure 2 It is the conditional diffusion decoding module.

[0098] Figure 3 This is a comparison of the method of the present invention with the benchmark model at CLIC2020. Detailed Implementation

[0099] This invention proposes a generative decoding method and system based on latent PSNR-adaptive diffusion time step and noise statistical matching. Under extremely low bit rate conditions, this scheme uses compressed latent data as a condition and adaptively determines the starting time step of diffusion backpropagation using the latent domain PSNR. Furthermore, it introduces noise statistical matching loss during the training phase to align the empirical statistics of sample differences with the theoretical diffusion noise statistics, thereby improving the perceived quality and stability of the reconstructed data while maintaining the bit rate (bpp).

[0100] Overall method flow:

[0101] S1 Conditional Acquisition and Bitrate Statistics

[0102] S1.1 Input Generation

[0103] To effectively capture the overall structural information of the image, this method first maps the input image to the VAE (Variational Auto-Encoder) latent space built into the diffusion model. This process is accomplished by the encoder. Extract global semantic features from the image while preserving key structural details such as edges and textures. Its mathematical form is as follows:

[0104] For the original image, For the diffusion model, an auto-encoder encoder, This represents the encoded raw global features.

[0105] S1.2 Feature Compression Module:

[0106] To further improve the efficiency of latent space coding, this method introduces the concept of Information Bottleneck (IB) in the feature compression module to perform structured compression of global features while maximizing the amount of information relevant to the reconstruction task. This module consists of a backbone latent encoder, a quantizer, and a decoder.

[0107] First of all Input with parameters encoder to obtain potential representations :

[0108]

[0109] After quantification Obtaining the latent representation :

[0110]

[0111] Finally, input the parameter. decoder get , representing the global features extracted from the decompression.

[0112]

[0113] To capture the spatial statistical relationships between latent features, this method further introduces a hyperprior structure. The parameter is A hyper-prior encoder. It is used to obtain edge information. To capture Spatial dependencies between elements:

[0114]

[0115] z is determined by the quantizer Quantification :

[0116]

[0117] Will enter with learning parameters decoder Generate entropy parameters .

[0118]

[0119] The advanced prior mechanism enhances the probabilistic modeling capability of latent representations, making compression more efficient and reconstruction more accurate.

[0120] S2 achieves time-step adaptation from the forward-diffusion-calibrated PSNR (PSNR → )

[0121] S2.1 PSNR–t calibration for forward diffusion

[0122] The forward propagation process of the diffusion model gradually injects Gaussian noise into the image. This method utilizes this gradual noise addition process to construct a monotonic correspondence between PSNR and time steps, thereby achieving "time step adaptation" in the inference stage. Given an original clean image... With diffusion schedule

[0123] Its forward (noise-adding) process is as follows:

[0124]

[0125] A set of noise images is generated step by step from 0 to T (where T can be set as needed, and in this patent, T=1000). .in Represents the given raw data , Indicates at time step Noisy data The distribution follows a mean of The standard deviation is The Gaussian distribution.

[0126] For each time step Calculate the MSE and PSNR of the latent domain compared to the original graph:

[0127]

[0128]

[0129] Where N is the total number of pixels. This is the image dynamic range constant. From this, a monotonically non-increasing calibration curve is obtained. :

[0130]

[0131] It can be stored as a lookup table and the corresponding time step can be found using the nearest neighbor interpolation method.

[0132] S2.2 During runtime, the time step is calculated by back-calculating the PSNR to transform the monotonic function of the above equation. Inverse equation:

[0133]

[0134] Based on this, the starting time step is obtained by estimating the image domain PSNR:

[0135]

[0136] In terms of implementation, The basic solution is achieved by combining a lookup table (LUT) with an inverse lookup, while satisfying the boundary conditions: when Time to take ;when Time to take .

[0137] S3 Noise Statistical Matching and Joint Optimization

[0138] Let the training objective be latent. This implementation uses the first four channels:

[0139]

[0140] in Represents the latent space feature tensor. This indicates that the tensor is a four-dimensional real tensor with a size of . B represents the number of input samples in a batch, C represents the number of feature channels, H represents the height of the feature map (pixels / grid size), and W represents the width of the feature map.

[0141] Calculate the mean squared error and PSNR over the latent domain:

[0142]

[0143]

[0144] The time step is obtained by performing a nearest neighbor lookup using the PSNR-t lookup table calibrated offline in S2.

[0145]

[0146] Subsequently The theoretical noise standard deviation was obtained. and variance .

[0147]

[0148]

[0149] Define nonnormalized latent residuals :

[0150]

[0151] Aggregate the tensors in spatial and channel dimensions, and then obtain scalar empirical statistics for all dimensions (b,c,h,w) in a batch:

[0152]

[0153] Joint optimization using noise statistical matching loss , From mean loss With variance loss Adding them together, we get:

[0154]

[0155]

[0156]

[0157] Image Reconstruction with S4 Diffusion Prior

[0158] To achieve high-perceptual-quality reconstruction results at extremely low bitrates, this invention employs a conditional diffusion decoding module. Without compromising existing generative capabilities, content variables are lightweightly injected into the diffusion denoising process, enabling the reconstructed image to possess both global structural consistency and realistic details.

[0159] S4.1 Overview of Stable Diffusion

[0160] First, using the first-stage encoder Image Encoding as a latent representation ,

[0161]

[0162] Then, a Markov chain is used to gradually increase the noise, with the noise intensity adjusted from the default noise parameters. Control, record

[0163]

[0164] This represents the data at diffusion step t. This represents pure noise sampled from a standard normal distribution. This is the cumulative noise retention term.

[0165] Noise estimator with U-Net structure learned during training To predict the added noise, the objective function is:

[0166]

[0167] This represents the loss function of the diffusion model. Represent the expected value of the following random variables: Clean images Text conditions: This patent is empty. For diffusion time step, It is random noise. For the model's output This represents the mean square error of the expected value.

[0168] S4.2 Conditional Diffusion Decoding Module

[0169] To fully utilize the generation prior of pre-trained stable diffusion under extremely low bitrate conditions, this invention performs lightweight conditional injection on the diffusion backbone.

[0170] S4.2.1 Overall Approach and Trainable Range

[0171] While maintaining the stability of the diffusion backbone, a small control branch is introduced to continuously inject conditional features related to the content variables into the diffusion model denoising network.

[0172] S4.2.2 Input and Channel Organization

[0173] The input to the control branch is a noise state. Conditional potential With text conditions It participates in the computation across layers by combining attention and temporal embeddings, and the results are returned to the sampler for use. (Corresponding:)

[0174]

[0175] S4.2.3 Structure and Injection Location

[0176] The Control Module is isomorphic to the Stable Diffusion backbone U-Net. At multiple layers, 1×1 zero-convolutions (Zero-Conv) are used to project the control module features onto the corresponding channels of the backbone and then inject them using residual summation. Three lists of zero-convolutions are maintained to achieve stable multi-point conditional fusion: layer-by-layer injection on the encoding side, single-point injection at bottlenecks, and layer-by-layer backfeeding on the decoding side.

[0177] S5 Reverse Diffusion Decoding Process

[0178] During the inference phase, an interval sampler is used to sample data at predetermined steps (steps = 50). By uniformly selecting locations from high noise to low noise and performing back-reasoning, the potential reconstruction can be obtained. Finally, the VAE decoder is used to... Reconstructed image :

[0179]

[0180] Joint Objective of S6 Training Phase

[0181] During the training phase, the following joint loss is used for end-to-end optimization:

[0182]

[0183] in:

[0184] The reconstruction loss is calculated as "L_1 / L_2" in the pixel domain and / or latent domain.

[0185] Perceived loss (which may include LPIPS, DISTS, etc.);

[0186] For rate distortion;

[0187] This is a noise statistics matching term.

[0188] This invention selects 84,991 high-quality training images from the LSDIR dataset and randomly crops them to a size of 512*512. In the experiments, we use StableDiffusion 2.1-base as the diffusion prior. For optimization, the Adam optimizer is used, with β1 parameter set to 0.9, β2 parameter set to 0.999, and the learning rate set to 1×10⁻⁻⁻⁶. 4 The training batch size was set to 4. We first trained the proposed model using λ=1 for 300,000 iterations; then we trained the model again using the target λ value for an additional 200,000 iterations. The learning rate during fine-tuning was set to 2×10⁻⁻⁻⁶. 5 .

[0189] Test data: The evaluation phase used 428 high-quality images from the CLIC2020 dataset. First, the image size was adjusted so that the shorter side was 768 pixels. Then, the images were center-cropped to obtain 768×768 resolution images for evaluation.

[0190] Evaluation Metrics: To conduct quantitative evaluation, several established metrics were employed to measure the perceptual quality of the results, including Learned Perceptual Patch Similarity (LPIPS), Natural Image Quality Evaluator (NIQE), Depth Image Structure and Texture Similarity (DISTS), Fréchet Inception Distance (FID), and Kernel Inception Distance (KID). Peak Signal-to-Noise Ratio (PSNR) and Multi-Scale Structural Similarity (MS-SSIM) were used to measure the fidelity of the reconstruction results. Furthermore, bitrate performance was evaluated using bits per pixel (bpp).

[0191] Figure 3 This paper demonstrates the performance of the training method of this invention compared to the benchmark model DiffEIC in terms of bit rate per pixel (bpp), learning-based perceptual patch similarity (LPIPS), natural image quality evaluator (NIQE), deep image structure and texture similarity (DISTS), Fréchet Inception distance (FID), and kernel Inception distance (KID). Simultaneously, we use peak signal-to-noise ratio (PSNR) and multi-scale structural similarity (MS-SSIM) to illustrate the performance. Here, bpp is calculated by dividing the base layer bit rate and the super-prior bit rate by the number of input pixels. As shown in the figure, the method of this invention achieves a significant improvement in performance such as PSNR.

Claims

1. An adaptive derivation method for diffusion time steps based on latent space PSNR, characterized in that, The method includes the following steps: S1 condition acquisition and bitrate statistics; S1.1 Input generation; First, the input image is mapped to the VAE latent space provided by the diffusion model; then, through the encoder... Extract global semantic features from the image while preserving edges and textures; its mathematical form is as follows: ； For the original image, For the diffusion model, an auto-encoder encoder, The encoded raw global features; S1.2 Feature compression module; The Information Bottleneck (IB) concept is introduced into the feature compression module to perform structured compression of global features while maximizing the amount of information relevant to the reconstruction task. The feature compression module consists of a backbone latent encoder, a quantizer, and a decoder. S2 achieves time-step adaptive PSNR from forward diffusion-calibrated PSNR → ; S2.1 PSNR–t calibration for forward diffusion; The forward propagation process of the diffusion model gradually injects Gaussian noise into the image; by utilizing this gradual noise addition process, a monotonic correspondence between PSNR and time steps is constructed, thereby achieving "time step adaptation" in the inference stage; given an original clean image... With diffusion schedule ; Its forward process is as follows: ； A set of noisy images is generated step by step from 0 to T. ;in Represents the given raw data , Indicates at time step Noisy data The distribution follows a mean of The standard deviation is Gaussian distribution; For each time step Calculate the MSE and PSNR of the latent domain compared to the original graph: ； ； Where N is the total number of pixels. The dynamic range constant of the image is used to obtain a monotonically non-increasing calibration curve. : ； It can be stored as a lookup table and the corresponding time step can be found using the nearest neighbor interpolation method; S2.2 During runtime, the time step is calculated by back-calculating the PSNR to transform the monotonic function of the above equation. Inverse equation: ； Based on this, the starting time step is obtained by estimating the image domain PSNR: ； The basic solution is achieved by combining a lookup table LUT with an inverse lookup, while satisfying the boundary conditions: when Time to take ;when Time to take ; S3 Noise Statistical Matching and Joint Optimization; Let the training objective be latent. Take the first 4 channels: ； in Represents the latent space feature tensor. This indicates that the tensor is a four-dimensional real tensor, B represents the number of input samples in a batch, C: the number of feature channels, H: the height of the feature map, and W: the width of the feature map; Calculate the mean squared error and PSNR over the latent domain: ； ； The time step is obtained by performing a nearest neighbor lookup using the PSNR-t lookup table calibrated offline in S2. ； Subsequently The theoretical noise standard deviation was obtained. and variance ; ； ； Define nonnormalized latent residuals : ； Aggregate the tensors in space and channel dimensions, and then obtain scalar empirical statistics for all dimensions (b,c,h,w) in a batch: ； Joint optimization using noise statistical matching loss , From mean loss With variance loss Adding them together, we get: ； ； ； Image reconstruction based on S4 diffusion prior; Use the conditional diffusion decoding module; Without compromising existing generative capabilities, content variables are injected into the diffusion denoising process in a lightweight manner, enabling the reconstructed image to have both global structural consistency and realistic details. S4.1 Overview of Stable Diffusion First, using the first-stage encoder Image Encoding as a latent representation , ； Then, a Markov chain is used to gradually increase the noise level, with the noise intensity adjusted from the default noise parameters. Control, record ； This represents the data at diffusion step t. This represents pure noise sampled from a standard normal distribution. This is the cumulative noise retention term; Noise estimator with U-Net structure learned during training To predict the added noise, the objective function is: ； This represents the loss function of the diffusion model. Represent the expected value of the following random variables: Clean image Text conditions For diffusion time step, It is random noise. For the model's output , representing the mean square error of the expected value; S4.2 Conditional Diffusion Decoding Module To fully utilize the generation prior of pre-trained stable diffusion under extremely low bitrate conditions, a lightweight conditional injection is performed on the diffusion backbone. S5 reverse diffusion decoding process; During the inference phase, an interval sampler is used to sample data at predetermined steps (steps = 50). By uniformly selecting locations from high noise to low noise and performing back-reasoning, the potential reconstruction can be obtained. Finally, the VAE decoder is used to... Reconstructed image : ； Joint objective of the S6 training phase; During the training phase, the following joint loss is used for end-to-end optimization: ； in: For reconstruction losses; To perceive loss; For rate distortion; This is a noise statistics matching term.

2. The adaptive derivation method for diffusion time step based on latent space PSNR according to claim 1, characterized in that, S1.2 includes: First, Input with parameters encoder to obtain potential representations : ； After quantification Obtaining the latent representation : ； Finally, input the parameter. decoder get , representing the global features extracted from the decompression; ； To capture the spatial statistical relationships between latent features, a hyperprior structure is introduced; The parameter is A hyper-prior encoder; used to obtain edge information. To capture Spatial dependencies between elements: ； z is determined by the quantizer Quantification : ； Will enter with learning parameters decoder Generate entropy parameters ; ； The advanced prior mechanism enhances the probabilistic modeling capability of latent representations, making compression more efficient and reconstruction more accurate.

3. The adaptive derivation method for diffusion time step based on latent space PSNR according to claim 1, characterized in that, S4.2 includes: S4.2.1 Overall Approach and Trainable Range: While maintaining the stable diffusion backbone capability, a small control branch is introduced to continuously inject conditional features related to content variables into the diffusion model denoising network; S4.2.2 Input and Channel Organization: The input to the control branch is a noise state. Conditional potential With text conditions By participating in the computation of each layer together with attention and temporal embedding, the results are returned to the sampler for use; S4.2.3 Structure and Injection Location: The Control Module is isomorphic to the Stable Diffusion backbone U-Net. At multiple layers, 1×1 zero-convolution Zero-Conv is used to project the control module features onto the channels corresponding to the backbone and then inject them by residual addition.

Images