Improving quantized performance of diffusion models via adaptive guidance
Adaptive guidance in diffusion models addresses 'cracking artifacts' by adjusting the guidance scale per time step, improving image quality and reducing latency in W8A8 quantization, achieving a 2x speedup with minimal overhead.
Patent Information
- Application Number
- PCT/CN2024/100006
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-06-19
- Publication Date
- 2025-12-26
AI Technical Summary
Existing post-training quantization (PTQ) and quantization-aware training (QAT) solutions for diffusion models result in substantial artifacts, known as 'cracking artifacts', when applying 8-bit weight (W8) 8-bit activation (A8) quantization (W8A8) for stable diffusion models, leading to increased latency and storage challenges.
Implement an adaptive guidance scale during quantization of diffusion models, adjusting the guidance scale per time step to combine conditional and unconditional latent spaces, using a decay function to minimize artifacts and improve image quality.
The adaptive guidance strategy significantly reduces 'cracking artifacts' and enhances signal quality noise ratio (SQNR) while maintaining efficient inference latency and storage requirements, enabling 2x speedup in inference speed with minimal hardware overhead.
Smart Images

Figure CN2024100006_26122025_PF_FP_ABST
Abstract
Description
IMPROVING QUANTIZED PERFORMANCE OF DIFFUSION MODELS VIA ADAPTIVE GUIDANCE
[0001] FIELD OF THE DISCLOSURE
[0002] Aspects of the present disclosure relate to artificial neural networks, and more specifically, to improving quantized performance of diffusion models.BACKGROUND
[0003] An artificial neural network, which may include an interconnected group of artificial neurons, may be a computational device or may represent a method to be performed by a computational device. Artificial neural networks may have corresponding structure and / or function in biological neural networks. Artificial neural networks, however, may provide useful computational techniques for certain applications, in which traditional computational techniques may be cumbersome, impractical, or inadequate. Because artificial neural networks may infer a function from observations, such networks may be useful in applications where the complexity of the task and / or data makes the design of the function burdensome using conventional techniques.
[0004] Image synthesis is a field of computer vision experiencing significant recent developments. Despite these recent significant developments, image synthesis involves substantial computational demands when performing high-resolution synthesis of complex, natural scenes. Recently, diffusion models have achieved impressive results in image synthesis. Diffusion models are probabilistic models designed to learn a data distribution by gradually denoising a normally distributed variable, which corresponds to learning the reverse process of a fixed Markov Chain. In practice, quantization is performed for diffusion models to reduce on-device inference latency. Unfortunately, existing post-training quantization (PTQ) and / or quantization-aware training (QAT) solutions result in substantial artifacts in generated images.SUMMARY
[0005] A processor-implemented method for adaptive guidance in diffusion models is described. The processor-implemented method includes identifying an inference step of a diffusion model. The processor-implemented method also includes determining an adaptive guidance scale value corresponding to the inference step of the diffusion model. The processor-implemented method further includes applying the adaptive guidance scale value to the inference step during quantization of the diffusion model.
[0006] An apparatus including at least one memory and at least one processor coupled to the at least one memory is described. The at least one processor is configured to identify an inference step of a diffusion model. The at least one processor is also configured to determine an adaptive guidance scale value corresponding to the inference step of the diffusion model. The at least one processor is further configured to apply the adaptive guidance scale value to the inference step during quantization of the diffusion model.
[0007] This has outlined, broadly, the features and technical advantages of the present disclosure in order that the detailed description that follows may be better understood. Additional features and advantages of the disclosure will be described below. It should be appreciated by those skilled in the art that this disclosure may be readily utilized as a basis for modifying or designing other structures for conducting the same purposes of the present disclosure. It should also be realized by those skilled in the art that such equivalent constructions do not depart from the teachings of the disclosure as set forth in the appended claims. The novel features, which are believed to be characteristic of the disclosure, both as to its organization and method of operation, together with further objects and advantages, will be better understood from the following description when considered in connection with the accompanying figures. It is to be expressly understood, however, that each of the figures is provided for the purpose of illustration and description only and is not intended as a definition of the limits of the present disclosure.BRIEF DESCRIPTION OF THE DRAWINGS
[0008] The features, nature, and advantages of the present disclosure will become more apparent from the detailed description set forth below when taken in conjunction with the drawings in which like reference characters identify correspondingly throughout.
[0009] FIGURE 1 illustrates an example implementation of a neural network using a system-on-a-chip (SOC) , including a general-purpose processor in accordance with certain aspects of the present disclosure.
[0010] FIGURES 2A, 2B, and 2C are diagrams illustrating a neural network in accordance with various aspects of the present disclosure.
[0011] FIGURE 2D is a diagram illustrating an exemplary deep convolutional network (DCN) in accordance with various aspects of the present disclosure.
[0012] FIGURE 3 is a block diagram illustrating a stable diffusion architecture, in accordance with various aspects of the present disclosure.
[0013] FIGURE 4 is a sample image generated based on a floating-point output.
[0014] FIGURES 5A and 5B illustrate generated images based on a compact representation, according to various aspects of the present disclosure.
[0015] FIGURES 6A and 6B illustrate generated images based on a compact representation utilizing an adaptive guidance scale, according to various aspects of the present disclosure.
[0016] FIGURES 7A-7C illustrate several types of adaptive guidance decay functions (f (t) ) graphs, according to various aspects of the present disclosure.
[0017] FIGURE 8 is a flow diagram illustrating an example processor-implemented method for adaptive guidance in diffusion models, in accordance with various aspects of the present disclosure.DETAILED DESCRIPTION
[0018] The detailed description set forth below, in connection with the appended drawings, is intended as a description of various configurations and is not intended to represent the only configurations in which the concepts described may be practiced. The detailed description includes specific details for the purpose of providing a thorough understanding of the various concepts. However, it will be apparent to those skilled in the art that these concepts may be practiced without these specific details. In some instances, well-known structures and components are shown in block diagram form to avoid obscuring such concepts.
[0019] Based on the teachings, one skilled in the art should appreciate that the scope of the disclosure is intended to cover any aspect of the disclosure, whether implemented independently of or combined with any other aspect of the disclosure. For example, an apparatus may be implemented or a method may be practiced using any number of the aspects set forth. In addition, the scope of the disclosure is intended to cover such an apparatus or method practiced using other structure, functionality, or structure and functionality in addition to or other than the various aspects of the disclosure set forth. Any aspect of the disclosure disclosed may be embodied by one or more elements of a claim.
[0020] The word “exemplary” is used to mean “serving as an example, instance, or illustration. ” Any aspect described as “exemplary” is not necessarily to be construed as preferred or advantageous over other aspects.
[0021] Although aspects are described, many variations and permutations of these aspects fall within the scope of the disclosure. Although some benefits and advantages of the preferred aspects are mentioned, the scope of the disclosure is not intended to be limited to benefits, uses or objectives. Rather, aspects of the disclosure are intended to be universally applicable to different technologies, system configurations, networks, and protocols, some of which are illustrated by way of example in the figures and in the following description of the preferred aspects. The detailed description and drawings are merely illustrative of the disclosure rather than limiting, the scope of the disclosure being defined by the appended claims and equivalents thereof.
[0022] Image synthesis is a field of computer vision having significant recent developments. Despite these recent significant developments, image synthesis involves substantial computational demands when performing high-resolution synthesis of complex, natural scenes. Recently, diffusion models have achieved impressive results in image synthesis. Diffusion models are probabilistic models designed to learn a data distribution by gradually denoising a normally distributed variable, which corresponds to learning the reverse process of a fixed Markov Chain. For image synthesis, models may rely on a reweighted variant of a variational lower bound on the learned data distribution. These models can be interpreted as an equally weighted sequence of denoising autoencoders that are trained to predict a denoised variant of their input.
[0023] In practice, quantization is performed for diffusion models to reduce on-device inference latency. For example, performing 8-bit weight (W8) 8-bit activation (A8) (W8A8) quantization for a stable diffusion model can offer significant benefits in terms of on-device inference latency. Currently, 8-bit weight (W8) 16-bit activation (A16) (W8A16) quantization is a popular solution for achieving stable diffusion quantization because the activation ranges of stable diffusion change iteration over iteration. This variation in the activation ranges makes storing activations in 8-bits highly challenging when using W8A8 quantization. Nevertheless, latency and storage specifications increase when using 16-bit activations relative to the reduced latency and storage specifications associated with 8-bit activations. Unfortunately, existing W8A8 post-training-quantization (PTQ) or quantization-aware-training (QAT) solutions result in substantial artifacts in generated images, which are referred to as cracking artifacts.
[0024] Various aspects of the present disclosure are directed to removing “cracking artifacts” in W8A8 (or other number of bits) model inference using an efficient inference time strategy referred to as adaptive guidance. According to various aspects of the present disclosure, a method for adaptive guidance in diffusion models includes identifying an inference step of a diffusion model. Once identified, the method includes generating an adaptive guidance scale corresponding to the inference step of the diffusion model. The method further includes applying the adaptive guidance scale to the inference step during quantization of the diffusion model.
[0025] FIGURE 1 illustrates an example implementation of a system-on-a-chip (SOC) 100, which may include a central processing unit (CPU) 102 or a multi-core CPU configured for adaptive guidance in diffusion models. Variables (e.g., neural signals and synaptic weights) , system parameters associated with a computational device (e.g., neural network with weights) , delays, frequency bin information, and task information may be stored in a memory block associated with a neural processing unit (NPU) 108, in a memory block associated with a CPU 102, in a memory block associated with a graphics processing unit (GPU) 104, in a memory block associated with a digital signal processor (DSP) 106, in a memory block 118, or may be distributed across multiple blocks. Instructions executed at the CPU 102 may be loaded from a program memory associated with the CPU 102 or may be loaded from a memory block 118.
[0026] The SOC 100 may also include additional processing blocks tailored to specific functions, such as a GPU 104, a DSP 106, a connectivity block 110, which may include fifth generation (5G) connectivity, fourth generation long term evolution (4G LTE) connectivity, Wi-Fi connectivity, USB connectivity, Bluetooth connectivity, and the like, and a multimedia processor 112 that may, for example, detect and recognize gestures. In one implementation, the NPU 108 is implemented in the CPU 102, DSP 106, and / or GPU 104. The SOC 100 may also include a sensor processor 114, image signal processors (ISPs) 116, and / or navigation module 120, which may include a global positioning system.
[0027] Each processor core of the multi-core CPU 102 may be a reduced instruction set computing (RISC) machine, RISC-V, an advanced RISC machine (ARM) , a microprocessor, or any reduced instruction set computing (RISC) architecture. The NPU 108 may be based on an ARM instruction set. The SOC 100 may be based on an ARM instruction set. Additionally, the NPU 108 may be based on an ARM instruction set. In an aspect of the present disclosure, the instructions loaded into the NPU 108 for adaptive guidance in diffusion models may include code to identify an inference step of a diffusion model. The instructions loaded into the NPU 108 may also include code to generate an adaptive guidance scale corresponding to the inference step of the diffusion model. The instructions loaded into the NPU 108 may further include code to apply the adaptive guidance scale to the inference step during quantization of the diffusion model.
[0028] Deep learning architectures may perform an object recognition task by learning to represent inputs at successively higher levels of abstraction in each layer, thereby building up a useful feature representation of the input data. In this way, deep learning addresses a major bottleneck of traditional machine learning. Prior to the advent of deep learning, a machine learning approach to an object recognition problem may have relied heavily on human engineered features, in combination with a shallow classifier. A shallow classifier may be a two-class linear classifier, for example, in which a weighted sum of the feature vector components may be compared with a threshold to predict to which class the input belongs. Human engineered features may be templates or kernels tailored to a specific problem domain by engineers with domain expertise. Deep learning architectures, in contrast, may learn to represent features that are like what a human engineer might design, but through training. Furthermore, a deep network may learn to represent and recognize new types of features that a human might not have considered.
[0029] A deep learning architecture may learn a hierarchy of features. If presented with visual data, for example, the first layer may learn to recognize simple features, such as edges, in the input stream. In another example, if presented with auditory data, the first layer may learn to recognize spectral power in specific frequencies. The second layer, taking the output of the first layer as input, may learn to recognize combinations of features, such as simple shapes for visual data or combinations of sounds for auditory data. For instance, higher layers may learn to represent complex shapes in visual data or words in auditory data. Still higher layers may learn to recognize common visual objects or spoken phrases.
[0030] Deep learning architectures may perform especially well when applied to problems that have a natural hierarchical structure. For example, the classification of motorized vehicles may benefit from first learning to recognize wheels, windshields, and other features. These features may be combined at higher layers in diverse ways to recognize cars, trucks, and airplanes.
[0031] Neural networks may be designed with a variety of connectivity patterns. In feed-forward networks, information is passed from lower to higher layers, with each neuron in each layer communicating to neurons in higher layers. A hierarchical representation may be built up in successive layers of a feed-forward network, as described above. Neural networks may also have recurrent or feedback (also called top-down) connections. In a recurrent connection, the output from a neuron in each layer may be communicated to another neuron in the same layer. A recurrent architecture may be helpful in recognizing patterns that span more than one of the input data chunks that are delivered to the neural network in a sequence. A connection from a neuron in each layer to a neuron in a lower layer is called a feedback (or top-down) connection. A network with many feedback connections may be helpful when the recognition of a high-level concept may aid in discriminating the low-level features of an input.
[0032] The connections between layers of a neural network may be fully connected or locally connected. FIGURE 2A illustrates an example of a fully connected neural network 202. In a fully connected neural network 202, a neuron in a first layer may communicate its output to every neuron in a second layer, so that each neuron in the second layer will receive input from every neuron in the first layer. FIGURE 2B illustrates an example of a locally connected neural network 204. In a locally connected neural network 204, a neuron in a first layer may be connected to a limited number of neurons in the second layer. More generally, a locally connected layer of the locally connected neural network 204 may be configured so that each neuron in a layer will have the same or a similar connectivity pattern, but with connections strengths that may have different values (e.g., 210, 212, 214, and 216) . The locally connected connectivity pattern may give rise to spatially distinct receptive fields in a higher layer because the higher layer neurons in each region may receive inputs that are tuned through training to the properties of a restricted portion of the total input to the network.
[0033] One example of a locally connected neural network is a convolutional neural network. FIGURE 2C illustrates an example of a convolutional neural network 206. The convolutional neural network 206 may be configured such that the connection strengths associated with the inputs for each neuron in the second layer are shared (e.g., 208) . Convolutional neural networks may be well suited to problems in which the spatial location of inputs is meaningful.
[0034] One type of convolutional neural network is a deep convolutional network (DCN) . FIGURE 2D illustrates a detailed example of a DCN 200 designed to recognize visual features from an image 226 input from an image capturing device 230, such as a car-mounted camera. The DCN 200 of the current example may be trained to identify traffic signs and a number provided on the traffic sign. Of course, the DCN 200 may be trained for other tasks, such as identifying lane markings or identifying traffic lights.
[0035] The DCN 200 may be trained with supervised learning. During training, the DCN 200 may be presented with an image, such as the image 226 of a speed limit sign, and a forward pass may then be computed to produce an output 222. The DCN 200 may include a feature extraction section and a classification section. Upon receiving the image 226, a convolutional layer 232 may apply convolutional kernels (not shown) to the image 226 to generate a first set of feature maps 218. As an example, the convolutional kernel for the convolutional layer 232 may be a 5x5 kernel that generates 28x28 feature maps. In the present example, because four different feature maps are generated in the first set of feature maps 218, four different convolutional kernels were applied to the image 226 at the convolutional layer 232. The convolutional kernels may also be referred to as filters or convolutional filters.
[0036] The first set of feature maps 218 may be subsampled by a max pooling layer (not shown) to generate a second set of feature maps 220. The max pooling layer reduces the size of the first set of feature maps 218. That is, a size of the second set of feature maps 220, such as 14x14, is less than the size of the first set of feature maps 218, such as 28x28. The reduced size provides similar information to a subsequent layer while reducing memory consumption. The second set of feature maps 220 may be further convolved via one or more subsequent convolutional layers (not shown) to generate one or more subsequent sets of feature maps (not shown) .
[0037] In the example of FIGURE 2D, the second set of feature maps 220 is convolved to generate a first feature vector 224. Furthermore, the first feature vector 224 is further convolved to generate a second feature vector 228. Each feature of the second feature vector 228 may include a number that corresponds to a feature of the image 226, such as “sign, ” “60, ” and “100. ” A softmax function (not shown) may convert the numbers in the second feature vector 228 to a probability. As such, an output 222 of the DCN 200 may be a probability of the image 226 including one or more features.
[0038] In the present example, the probabilities in the output 222 for “sign” and “60” are higher than the probabilities of the others of the output 222, such as “30, ” “40, ” “50, ” “70, ” “80, ” “90, ” and “100” . Before training, the output 222 produced by the DCN 200 may be incorrect. Thus, an error may be calculated between the output 222 and a target output. The target output is the ground truth of the image 226 (e.g., “sign” and “60” ) . The weights of the DCN 200 may then be adjusted so the output 222 of the DCN 200 is more closely aligned with the target output.
[0039] To adjust the weights, a learning algorithm may compute a gradient vector for the weights. The gradient may indicate an amount that an error would increase or decrease if the weight were adjusted. At the top layer, the gradient may correspond directly to the value of a weight connecting an activated neuron in the penultimate layer and a neuron in the output layer. In lower layers, the gradient may depend on the value of the weights and on the computed error gradients of the higher layers. The weights may then be adjusted to reduce the error. This manner of adjusting the weights may be referred to as “back propagation” as it involves a “backward pass” through the neural network.
[0040] In practice, the error gradient of weights may be calculated over a small number of examples, so that the calculated gradient approximates the true error gradient. This approximation method may be referred to as stochastic gradient descent. Stochastic gradient descent may be repeated until the achievable error rate of the entire system has stopped decreasing or until the error rate has reached a target level. After learning, the DCN 200 may be presented with new images (e.g., the speed limit sign of the image 226 or another speed limit sign or other image) and a forward pass through the DCN 200 may yield an output 222 that may be considered an inference or a prediction of the DCN 200.
[0041] Deep belief networks (DBNs) are probabilistic models comprising multiple layers of hidden nodes. DBNs may be used to extract a hierarchical representation of training data sets. A DBN may be obtained by stacking up layers of Restricted Boltzmann Machines (RBMs) . An RBM is a type of artificial neural network that can learn a probability distribution over a set of inputs. Because RBMs can learn a probability distribution in the absence of information about the class to which each input should be categorized, RBMs are often used in unsupervised learning. Using a hybrid unsupervised and supervised paradigm, the bottom RBMs of a DBN may be trained in an unsupervised manner and may serve as feature extractors, and the top RBM may be trained in a supervised manner (on a joint distribution of inputs from the previous layer and target classes) and may serve as a classifier.
[0042] DCNs are networks of convolutional networks, configured with additional pooling and normalization layers. DCNs have achieved state-of-the-art performance on many tasks. DCNs can be trained using supervised learning in which both the input and output targets are known for many exemplars and are used to modify the weights of the network by use of gradient descent methods.
[0043] DCNs may be feed-forward networks. In addition, as described above, the connections from a neuron in a first layer of a DCN to a group of neurons in the next higher layer are shared across the neurons in the first layer. The feed-forward and shared connections of DCNs may be exploited for fast processing. The computational burden of a DCN may be much less, for example, than that of a similarly sized neural network that comprises recurrent or feedback connections.
[0044] The processing of each layer of a convolutional network may be considered a spatially invariant template or basis projection. If the input is first decomposed into multiple channels, such as the red, green, and blue channels of a color image, then the convolutional network trained on that input may be considered three-dimensional, with two spatial dimensions along the axes of the image and a third dimension capturing color information. The outputs of the convolutional connections may be considered to form a feature map in the subsequent layer, with each element of the feature map (e.g., 220) receiving input from a range of neurons in the previous layer (e.g., feature maps 218) and from each of the multiple channels. The values in the feature map may be further processed with a non-linearity, such as a rectification, max (0, x) . Values from adjacent neurons may be further pooled, which corresponds to down sampling, and may provide additional local invariance and dimensionality reduction. Normalization, which corresponds to whitening, may also be applied through lateral inhibition between neurons in the feature map.
[0045] Image synthesis is a field of computer vision having significant recent developments. Recently, diffusion models have achieved impressive results in image synthesis. Diffusion models are probabilistic models designed to learn a data distribution by gradually denoising a normally distributed variable, which corresponds to learning the reverse process of a fixed Markov Chain. For image synthesis, models may rely on a reweighted variant of a variational lower bound on the learned data distribution. These models can be interpreted as an equally weighted sequence of denoising autoencoders that are trained to predict a denoised variant of their input.
[0046] In practice, quantization is performed for diffusion models to reduce on-device inference latency. For example, performing 8-bit weight (W8) 8-bit activation (A8) (W8A8) quantization for a stable diffusion model can offer significant benefits in terms of on-device inference latency. Currently, 8-bit weight (W8) 16-bit activation (A16) (W8A16) quantization is a popular solution for achieving stable diffusion quantization because the activation ranges of stable diffusion change iteration over iteration. This variation in the activation ranges makes storing activations in 8-bits highly challenging when using W8A8 quantization. Nevertheless, latency and storage specifications increase when using 16-bit activations relative to the reduced latency and storage specifications associated with 8-bit activations. Unfortunately, existing W8A8 post-training-quantization (PTQ) or quantization-aware-training (QAT) solutions result in substantial artifacts in generated images, which are referred to as cracking artifacts.
[0047] Various aspects of the present disclosure are directed to removing “cracking artifacts” in W8A8 (or other number of bits) model inference using an efficient inference time strategy referred to as adaptive guidance. According to various aspects of the present disclosure, a method for adaptive guidance in diffusion models includes identifying an inference step of a diffusion model. Once identified, the method includes generating an adaptive guidance scale corresponding to the inference step of the diffusion model. The method further includes applying the adaptive guidance scale to the inference step during quantization of the diffusion model. A stable diffusion architecture is shown in FIGURE 3. Although described with reference to stable diffusion, it should be recognized that the disclosed adaptive guidance may be applied to quantization of other diffusion models, or other like neural network models.
[0048] FIGURE 3 is a block diagram illustrating a stable diffusion architecture, in accordance with various aspects of the present disclosure. As shown in FIGURE 3, a stable diffusion architecture 300 includes a pixel space 310, a latent space 320, and a conditioning block 330. In this example, the pixel space 310 includes an encoder ε, which encodes an image x into a latent representation z = ε (x) , and a decoder which reconstructs the image from the latent, giving Additionally, the latent space 320 includes a diffusion process to generate a latent Z at time T (ZT) , a denoising U-Net εθ 340, and a denoising block 350. In this example, the conditioning block 330 includes a semantic map, text, representations, and / or images, which are encoded by an encoder τθ and supplied as a concatenation 332.
[0049] As shown in FIGURE 3, the denoising U-Net εθ 340 receives the latent ZT and a switch 342 determines whether the concatenation 332 is fed to the U-Net εθ 340 and the denoising block 350 for computation of a conditional or unconditional latent space. The denoising U-Net εθ 340 further includes an encoder and a decoder having cross-attention blocks 360, in which skip connections are shown. An output of the denoising U-Net εθ 340 is a previous latent ZT-1, which is provided to the denoising block 350 to generate the latent Z, which is provided to the pixel space 310.
[0050] In this example, the denoising U-Net εθ 340 illustrates an inference strategy at each time step (t) , which is computed according to Equation (1)
[0051] Latents (t) = unconditional latent (t) + guidance scale × [conditional latent (t) –unconditional latent (t) ] (1)
[0052] As shown in Equation (1) , a fixed guidance scale is applied during inference to combine a conditional latent space and unconditional latent space of the stable diffusion architecture, according to a setting of the switch 342. In practice, diffusion models use a fixed guidance scale during inference prediction to combine conditional and unconditional latent spaces of the generative model for each time step t, as shown in Equation (1) . This approach may result in generated images having a degraded image quality, including artifacts, particularly when configured with low bit-width quantized models (e.g., a W8A8 quantized diffusion model) . Although described with reference to W8A8 quantization, it should be recognized that aspects of the present disclosure are not limited to W8A8 quantization and may be applied to WXAY quantization for X-bit weights and Y-bit activations. As described, the artifacts arising in W8A8 quantized models are referred to as “cracking artifacts” , for example, as shown in FIGURES 5A and 5B.
[0053] FIGURE 4 is a sample image generated based on a floating-point output. In this example, an image 400 is generated based on the following caption: “Beautiful sunset by the sea and mountain. ” As shown in FIGURE 4, the image 400 is generated based on a floating-point output, rather than an integer 8-bit (INT8) representation, as shown in FIGURES 5A-6B, relative to the image 400 generated based on a floating-point output.
[0054] FIGURES 5A and 5B illustrate generated images based on a compact representation, according to various aspects of the present disclosure. As shown in FIGURES 5A and 5B, images are generated based on the caption: “Beautiful sunset by the sea and mountain, ” as used to generate the image 400 of FIGURE 4. As shown in FIGURE 5A, a first image 500 is generated based on a W8A8 PTQ fixed guidance output using Equation (1) . The first image 500, relative to the image 400 of FIGURE 4, exhibits clear cracking artifacts and a reduced signal quality noise ratio (SQNR) (e.g., 4.7 dB) . As shown in FIGURE 5B, a second image 550 is generated based on a W8A8 QAT fixed guidance output using Equation (1) . The second image 550, relative to the image 400 of FIGURE 4, exhibits clear cracking artifacts and a reduced SQNR (e.g., 6.1 dB) .
[0055] Various aspects of the present disclosure are directed to removing “cracking artifacts” appearing in W8A8 (or other number of bits) model inferences shown in FIGURES 5A and 5B through the utilization of an efficient inference time strategy referred to as adaptive guidance. According to various aspects of the present disclosure, a method for adaptive guidance in diffusion models includes utilization of adaptive guidance. In various aspects of the present disclosure, an adaptive guidance scale is applied during a combining of conditional latent space and unconditional latent space of the inference step of the diffusion model. In these aspects of the present disclosure, an adaptive guidance scale function G (t) may be utilized to generate a guidance scale for each time step of inference to improve the generational capabilities of the quantized diffusion model (e.g., W8A8 PTQ / W8A8 QAT) , according to Equation (2) : Latents (t) = unconditional latent (t) + G (t) × [conditional latent (t) –unconditional latent (t) ] (2)
[0056] FIGURES 6A and 6B illustrate generated images based on a compact representation utilizing an adaptive guidance scale, according to various aspects of the present disclosure. As shown in FIGURES 6A and 6B, images are generated based on the caption: “Beautiful sunset by the sea and mountain, ” as used to generate the image 400 of FIGURE 4 and the images 500 and 550 of FIGURES 5A and 5B. As shown in FIGURE 6A, a first image 600 is generated based on a W8A8 PTQ adaptive guidance output using Equation (2) . The first image 600, relative to the first image 500 of FIGURE 5A, reduces the clear cracking artifacts and exhibits an improved SQNR (e.g., 7.0 dB) . As shown in FIGURE 6B, a second image 650 is generated based on a W8A8 QAT adaptive guidance output using Equation (2) . The second image 650, relative to the second image 550 of FIGURE 4, further reduces the clear cracking artifacts and exhibits a further improved SQNR (e.g., 7.8 dB) .
[0057] As in Equation (2) , the adaptive guidance scale may be implemented utilizing a function G (t) that decays with time to significantly reduce cracking artifacts. As shown in Equation (3) , for a start and an end guidance scale, λ and β, respectively, an adaptive guidance scale can be modeled as: G (t) = β + (λ–β) f (t) (3)
[0058] where f (t) , the decay function, satisfies f (t+1) < f (t) . In some examples, f (0) = 1 (no decay at the start) , and / or lim (t → Tfinal) f (t) = 0 (complete decay over time) , but the decay function need not start at one or decay to zero in all implementations or situations. In some examples, λ and / or β also vary over time. As can be recognized based on the above description, the function G (t) , which may be used to implement the adaptive guidance scale, may satisfy G (t+1) < G (t) . Additionally, values in the adaptive guidance scale may be determined in any number of manners. For example, values in the adaptive guidance scale may be calculated or generated based on a function (e.g., G(t) ) , could be loaded from a preexisting scale (for example, from memory or registers) , etc.
[0059] FIGURES 7A-7C illustrate several types of adaptive guidance decay functions (f (t) ) graphs, according to various aspects of the present disclosure. In particular, the examples shown in FIGURES 7A-7C are based on adaptive guidance with different decay with a start guidance value, λ=10.5, and various end guidance (β) values. For example, FIGURE 7A illustrates an adaptive guidance linear decay function graph 700. Additionally, FIGURE 7B illustrates an adaptive guidance exponential decay function graph 730. FIGURE 7C illustrates an adaptive guidance cosine decay function graph 760. For example, FIGURES 6A and 6B illustrate images generated based on linear time decay adaptive guidance (e.g., λ = 10, β = 2.5) .
[0060] According to various aspects of the present disclosure, the adaptive guidance scheme improves the low-bit quantized results by significantly removing artifacts from a generated output image. Beneficially, the adaptive guidance scheme involves negligible hardware deployment overhead. Additionally, enabling W8A8 quantization for diffusion models (e.g., stable diffusion) can improve inference latency on-device. In particular, the W8A8 quantization described above results in a theoretical inference speed up (e.g., 2x) with minimal overhead.
[0061] Various aspects of the present disclosure, however, recognize that sensitivity to quantization increase with time steps (t) . For example, a dynamic range of a particular sensitive activation (e.g., SoftMax) increases with time steps (t) . Thus, using the same quantization parameters (e.g., bit-width) for all time steps can cause degradation, for example for certain time steps with smaller dynamic ranges. Accordingly, various aspects of the present disclosure provide a time-step aware mixed precision to enable low bit quantization of diffusion models, in which the adaptive guidance scale of Equation (2) is varied per time step. Additionally, various aspects of the present disclosure provide a time step aware, quantization bit-width allocation. As recognized by aspects of the present disclosure, sensitivity to quantization increases with time steps for 8-bit quantization due to degradation caused by quantization noise.
[0062] Various aspects of the present disclosure perform a time step aware, quantization bit-width allocation for providing higher precision in later time steps to solve the issue of degradation due to the quantization noise. For example, using a higher bit width (e.g., more than W8A8) in later timesteps reduces degradation. In some examples, the bit width increases after a certain number of time steps, or increases when a certain dynamic range or noise threshold is reached, although other criteria for determining when to increase the bit width are possible. A process for adaptive guidance is described, for example, in FIGURE 8.
[0063] FIGURE 8 is a flow diagram illustrating an example processor-implemented method 800 for adaptive guidance in diffusion models, in accordance with various aspects of the present disclosure. The method 800 begins at block 802, in which an inference step of a diffusion model is identified. At block 804, an adaptive guidance scale value is determined corresponding to the inference step of the diffusion model. At block 806, the adaptive guidance scale value is applied to the inference step during quantization of the diffusion model. For example, a function G (t) is utilized to generate a guidance scale for each time step of inference to improve the generational capabilities of the quantized diffusion model (e.g., W8A8 PTQ / W8A8 QAT) , according to Equation (2) , as shown above.
[0064] Implementation examples are described in the following numbered clauses:
[0065] 1. A processor-implemented method for adaptive guidance in diffusion models, comprising:
[0066] identifying an inference step of a diffusion model;
[0067] determining an adaptive guidance scale value corresponding to the inference step of the diffusion model; and
[0068] applying the adaptive guidance scale value to the inference step during quantization of the diffusion model.
[0069] 2. The processor-implemented method of clause 1, in which the adaptive guidance scale value varies per time step.
[0070] 3. The processor-implemented method of any of clauses 1 or 2, in which a quantization bit-width varies per time step.
[0071] 4. The processor-implemented method of any of clauses 1-3, in which the applying is performed during post-training quantization (PTQ) .
[0072] 5. The processor-implemented method of any of clauses 1-3, in which the applying is performed during quantization-aware training (QAT) .
[0073] 6. The processor-implemented method of any of clauses 1-5, in which the diffusion model comprises a stable diffusion model.
[0074] 7. The processor-implemented method of any of clauses 1-6, in which the adaptive guidance scale value is applied during a combining of conditional latent space and unconditional latent space of the inference step of the diffusion model.
[0075] 8. The processor-implemented method of any of clauses 1-7, in which the adaptive guidance scale value is configured to decay with time.
[0076] 9. The processor-implemented method of any of clauses 1-7, in which the adaptive guidance scale value is configured to decay with linear decay.
[0077] 10. The processor-implemented method of any of clauses 1-7, in which the adaptive guidance scale value is configured to decay with exponential decay.
[0078] 11. The processor-implemented method of any of clauses 1-7, in which the adaptive guidance scale value is configured to decay with cosine decay.
[0079] 12. The processor-implemented method of any of clauses 1-11, in which the quantization comprises 8-bit weight (W8) 8-bit activation (A8) (W8A8) quantization.
[0080] 13. An apparatus, comprising:
[0081] at least one memory; and
[0082] at least one processor coupled to the at least one memory, the at least one processor configured to:
[0083] identify an inference step of a diffusion model;
[0084] determine an adaptive guidance scale value corresponding to the inference step of the diffusion model; and
[0085] apply the adaptive guidance scale value to the inference step during quantization of the diffusion model.
[0086] 14. The apparatus of clause 13, in which the adaptive guidance scale value varies per time step.
[0087] 15. The apparatus of clause 13, in which a quantization bit-width varies per time step.
[0088] 16. The apparatus of any of clauses 13-15, in which the processor is further configured to apply the adaptive guidance scale value during post-training quantization (PTQ) .
[0089] 17. The apparatus of any of clauses 13-15, in which the processor is further configured to apply the adaptive guidance scale value during quantization-aware training (QAT) .
[0090] 18. The apparatus of any of clauses 13-15, in which the diffusion model comprises a stable diffusion model.
[0091] 19. The apparatus of any of clauses 13-18, in which the processor is further configured to apply the adaptive guidance scale value during a combining of conditional latent space and unconditional latent space of the inference step of the diffusion model.
[0092] 20. The apparatus of any of clauses 13-19, in which the adaptive guidance scale value is configured to decay with time.
[0093] 21. The apparatus of any of clauses 13-19, in which the adaptive guidance scale value is configured to decay with linear decay.
[0094] 22. The apparatus of any of clauses 13-19, in which the adaptive guidance scale value is configured to decay with exponential decay.
[0095] 23. The apparatus of any of clauses 13-19, in which the adaptive guidance scale value is configured to decay with cosine decay.
[0096] 24. The apparatus of any of clauses 13-23, in which the quantization comprises 8-bit weight (W8) 8-bit activation (A8) (W8A8) quantization.
[0097] The various operations of methods described above may be performed by any suitable means capable of performing the corresponding functions. The means may include various hardware and / or software component (s) and / or module (s) , including, but not limited to, a circuit, an application specific integrated circuit (ASIC) , or processor. Where there are operations illustrated in the figures, those operations may have corresponding counterpart means-plus-function components with similar numbering.
[0098] As used, the term “determining” encompasses a wide variety of actions. For example, “determining” may include calculating, computing, processing, deriving, investigating, looking up (e.g., looking up in a table, a database, or another data structure) , ascertaining and the like. Additionally, “determining” may include receiving (e.g., receiving information) , accessing (e.g., accessing data in a memory) and the like. Furthermore, “determining” may include resolving, selecting, choosing, establishing, and the like.
[0099] As used, a phrase referring to “at least one of” a list of items refers to any combination of those items, including single members. As an example, “at least one of: a, b, or c” is intended to cover: a, b, c, a-b, a-c, b-c, and a-b-c.
[0100] The various illustrative logical blocks, modules and circuits described in connection with the present disclosure may be implemented or performed with a general-purpose processor, a digital signal processor (DSP) , an application specific integrated circuit (ASIC) , a field programmable gate array signal (FPGA) or other programmable logic device (PLD) , discrete gate or transistor logic, discrete hardware components or any combination thereof designed to perform the functions described. A general-purpose processor may be a microprocessor, but in the alternative, the processor may be any commercially available processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration.
[0101] The steps of a method or algorithm described in connection with the present disclosure may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module may reside in any form of storage medium that is known in the art. Some examples of storage media that may be used include random access memory (RAM) , read only memory (ROM) , flash memory, erasable programmable read-only memory (EPROM) , electrically erasable programmable read-only memory (EEPROM) , registers, a hard disk, a removable disk, a CD-ROM and so forth. A software module may comprise a single instruction, or many instructions, and may be distributed over several different code segments, among different programs, and across multiple storage media. A storage medium may be coupled to a processor such that the processor can read information from, and write information to, the storage medium. In the alternative, the storage medium may be integral to the processor.
[0102] The methods disclosed comprise one or more steps or actions for achieving the described method. The method steps and / or actions may be interchanged with one another without departing from the scope of the claims. In other words, unless a specific order of steps or actions is specified, the order and / or use of specific steps and / or actions may be modified without departing from the scope of the claims.
[0103] The functions described may be implemented in hardware, software, firmware, or any combination thereof. If implemented in hardware, an example hardware configuration may comprise a processing system in a device. The processing system may be implemented with a bus architecture. The bus may include any number of interconnecting buses and bridges depending on the specific application of the processing system and the overall design constraints. The bus may link together various circuits including a processor, machine-readable media, and a bus interface. The bus interface may be used to connect a network adapter, among other things, to the processing system via the bus. The network adapter may be used to implement signal processing functions. For certain aspects, a user interface (e.g., keypad, display, mouse, joystick, etc. ) may also be connected to the bus. The bus may also link various other circuits such as timing sources, peripherals, voltage regulators, power management circuits, and the like, which are well known in the art, and therefore, will not be described any further.
[0104] The processor may be responsible for managing the bus and general processing, including the execution of software stored on the machine-readable media. The processor may be implemented with one or more general-purpose and / or special-purpose processors. Examples include microprocessors, microcontrollers, DSP processors, and other circuitry that can execute software. Software shall be construed broadly to mean instructions, data, or any combination thereof, whether referred to as software, firmware, middleware, microcode, hardware description language, or otherwise. Machine-readable media may include, by way of example, random access memory (RAM) , flash memory, read only memory (ROM) , programmable read-only memory (PROM) , erasable programmable read-only memory (EPROM) , electrically erasable programmable Read-only memory (EEPROM) , registers, magnetic disks, optical disks, hard drives, or any other suitable storage medium, or any combination thereof. The machine-readable media may be embodied in a computer-program product. The computer-program product may comprise packaging materials.
[0105] In a hardware implementation, the machine-readable media may be part of the processing system separate from the processor. However, as those skilled in the art will readily appreciate, the machine-readable media, or any portion thereof, may be external to the processing system. By way of example, the machine-readable media may include a transmission line, a carrier wave modulated by data, and / or a computer product separate from the device, all which may be accessed by the processor through the bus interface. Alternatively, or in addition, the machine-readable media, or any portion thereof, may be integrated into the processor, such with cache and / or general register files. Although the various components discussed may be described as having a specific location, such as a local component, they may also be configured in numerous ways, such as certain components being configured as part of a distributed computing system.
[0106] The processing system may be configured as a general-purpose processing system with one or more microprocessors providing the processor functionality and external memory providing at least a portion of the machine-readable media, all linked together with other supporting circuitry through an external bus architecture. Alternatively, the processing system may comprise one or more neuromorphic processors for implementing the neuron models and models of neural systems described. As another alternative, the processing system may be implemented with an application specific integrated circuit (ASIC) with the processor, the bus interface, the user interface, supporting circuitry, and at least a portion of the machine-readable media integrated into a single chip, or with one or more field programmable gate arrays (FPGAs) , programmable logic devices (PLDs) , controllers, state machines, gated logic, discrete hardware components, or any other suitable circuitry, or any combination of circuits that can perform the various functionality described throughout this disclosure. Those skilled in the art will recognize how best to implement the described functionality for the processing system depending on the application and the overall design constraints imposed on the overall system.
[0107] The machine-readable media may comprise several software modules. The software modules include instructions that, when executed by the processor, cause the processing system to perform various functions. The software modules may include a transmission module and a receiving module. Each software module may reside in a single storage device or be distributed across multiple storage devices. By way of example, a software module may be loaded into RAM from a hard drive when a triggering event occurs. During execution of the software module, the processor may load some of the instructions into cache to increase access speed. One or more cache lines may then be loaded into a general register file for execution by the processor. When referring to the functionality of a software module below, it will be understood that such functionality is implemented by the processor when executing instructions from that software module. Furthermore, it should be appreciated that aspects of the present disclosure result in improvements to the functioning of the processor, computer, machine, or other system implementing such aspects.
[0108] If implemented in software, the functions may be stored or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media include both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A storage medium may be any available medium that can be accessed by a computer. By way of example, and not limitation, such computer-readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. Additionally, any connection is properly termed a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL) , or wireless technologies such as infrared (IR) , radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of medium. Disk and disc, as used, include compact disc (CD) , laser disc, optical disc, digital versatile disc (DVD) , floppy disk, and disc where disks usually reproduce data magnetically, while discs reproduce data optically with lasers. Thus, in some aspects, computer-readable media may comprise non-transitory computer-readable media (e.g., tangible media) . In addition, for other aspects computer-readable media may comprise transitory computer-readable media (e.g., a signal) . Combinations of the above should also be included within the scope of computer-readable media.
[0109] Thus, certain aspects may comprise a computer program product for performing the operations presented. For example, such a computer program product may comprise a computer-readable medium having instructions stored (and / or encoded) thereon, the instructions being executable by one or more processors to perform the operations described. For certain aspects, the computer program product may include packaging material.
[0110] Further, it should be appreciated that modules and / or other appropriate means for performing the methods and techniques described can be downloaded and / or otherwise obtained by a user terminal and / or base station as applicable. For example, such a device can be coupled to a server to facilitate the transfer of means for performing the methods described. Alternatively, various methods described can be provided via storage means (e.g., RAM, ROM, a physical storage medium such as a compact disc (CD) or floppy disk, etc. ) , such that a user terminal and / or base station can obtain the various methods upon coupling or providing the storage means to the device. Moreover, any other suitable technique for providing the methods and techniques described to a device can be utilized.
[0111] It is to be understood that the claims are not limited to the precise configuration and components illustrated above. Various modifications, changes, and variations may be made in the arrangement, operation, and details of the methods and apparatus described above without departing from the scope of the claims.
Claims
1.A processor-implemented method for adaptive guidance in diffusion models, comprising:identifying an inference step of a diffusion model;determining an adaptive guidance scale value corresponding to the inference step of the diffusion model; andapplying the adaptive guidance scale value to the inference step during quantization of the diffusion model.2.The processor-implemented method of claim 1, in which the adaptive guidance scale value varies per time step.3.The processor-implemented method of claim 1, in which a quantization bit-width varies per time step.4.The processor-implemented method of claim 1, in which the applying is performed during post-training quantization (PTQ) .5.The processor-implemented method of claim 1, in which the applying is performed during quantization-aware training (QAT) .6.The processor-implemented method of claim 1, in which the diffusion model comprises a stable diffusion model.7.The processor-implemented method of claim 1, in which the adaptive guidance scale value is applied during a combining of conditional latent space and unconditional latent space of the inference step of the diffusion model.8.The processor-implemented method of claim 1, in which the adaptive guidance scale value is configured to decay with time.9.The processor-implemented method of claim 1, in which the adaptive guidance scale value is configured to decay with linear decay.10.The processor-implemented method of claim 1, in which the adaptive guidance scale value is configured to decay with exponential decay.11.The processor-implemented method of claim 1, in which the adaptive guidance scale value is configured to decay with cosine decay.12.The processor-implemented method of claim 1, in which the quantization comprises 8-bit weight (W8) 8-bit activation (A8) (W8A8) quantization.13.An apparatus, comprising:at least one memory; andat least one processor coupled to the at least one memory, the at least one processor configured to:identify an inference step of a diffusion model;determine an adaptive guidance scale value corresponding to the inference step of the diffusion model; andapply the adaptive guidance scale value to the inference step during quantization of the diffusion model.14.The apparatus of claim 13, in which the adaptive guidance scale value varies per time step.15.The apparatus of claim 13, in which a quantization bit-width varies per time step.16.The apparatus of claim 13, in which the processor is further configured to apply the adaptive guidance scale value during post-training quantization (PTQ) .17.The apparatus of claim 13, in which the processor is further configured to apply the adaptive guidance scale value during quantization-aware training (QAT) .18.The apparatus of claim 13, in which the diffusion model comprises a stable diffusion model.19.The apparatus of claim 13, in which the processor is further configured to apply the adaptive guidance scale value during a combining of conditional latent space and unconditional latent space of the inference step of the diffusion model.20.The apparatus of claim 13, in which the adaptive guidance scale value is configured to decay with time.
Citation Information
Patent Citations
Image generation method and device, equipment and medium
CN116701692A
Method and device for encoding and decoding image generated by artificial intelligence model
CN116886937A
Method for correction of artifacts arising from temporal changes of attenuation values
US20120051619A1
Layer-level quantization in neural networks
US20190171927A1
Null-text inversion for editing real images using guided diffusion models
WO2024107884A1