{
  "id": "latent-diffusion-model",
  "title": "Latent diffusion model",
  "lang": "en",
  "source_lang": "en",
  "summary": "A generative model that learns to reverse noise corruption in a compressed representation, enabling efficient synthesis and editing of images and other data.",
  "category": "computer-science",
  "subcategory": "artificial-intelligence",
  "type": "concept",
  "aliases": [
    "Latent diffusion",
    "LDM"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/latent-diffusion-model/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/latent-diffusion-model/",
      "json": "https://aiwiki.page/zh/latent-diffusion-model.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A latent diffusion model is a generative artificial intelligence model that synthesizes data by progressively removing noise from a learned, compressed representation rather than directly from the original data. In image generation, an encoder first maps an image into a lower-resolution latent space. A diffusion model learns the statistical structure of these representations, and a decoder converts a generated representation back into an image. Text, images, spatial layouts, or other information can guide the generation process.\n\nThe approach separates two demanding tasks: representing perceptually important information and learning how to generate new examples. Performing the iterative generation process in a smaller representation substantially reduces its computational cost compared with otherwise similar pixel-space diffusion systems. Latent diffusion became especially influential through Stable Diffusion, publicly released in 2022, and subsequently supported image editing, personalization, video generation, and audio synthesis. Its capabilities depend jointly on the compression model, the denoising network, the conditioning mechanism, the training data, and the numerical procedure used for generation."
    },
    {
      "heading": "Historical foundations and emergence",
      "text": "Latent diffusion combines developments in generative modeling, learned compression, and iterative denoising. Autoencoders established an encoder–decoder framework in which observations are mapped to an internal representation and reconstructed from it. Variational autoencoders, introduced in work by Diederik Kingma and Max Welling in 2013 and independently developed in related work by Danilo Rezende and colleagues, added a probabilistic formulation and regularized latent representations. These systems demonstrated that a learned intermediate space could support both reconstruction and generation, although their objectives could produce visibly smooth reconstructions.\n\nA separate line of work developed diffusion-based generative modeling. Jascha Sohl-Dickstein and colleagues introduced a framework in 2015 that gradually corrupts data through a forward diffusion process and learns a reverse process. Jonathan Ho, Ajay Jain, and Pieter Abbeel’s 2020 paper on denoising diffusion probabilistic models demonstrated strong image synthesis with a simplified denoising objective. Related score-based methods developed by Yang Song and collaborators connected generation to estimates of the gradient of a data distribution’s logarithmic density. These approaches made iterative denoising a competitive alternative to generative adversarial networks.\n\nDirect diffusion over images, however, repeatedly processes large spatial arrays. A high-resolution image contains many values that describe fine local variations rather than broad semantic structure. Modeling all of these values during every denoising step creates a substantial computational burden. Earlier latent-variable generators already used learned representations, but latent diffusion made the combination of perceptual compression and a diffusion prior particularly effective for high-resolution synthesis.\n\nRobin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer presented High-Resolution Image Synthesis with Latent Diffusion Models as a preprint in December 2021 and at the 2022 Conference on Computer Vision and Pattern Recognition. Their system trained an image autoencoder first and then trained diffusion models on its representations. The work explored different compression factors and conditioning tasks, including text-to-image synthesis, inpainting, and super-resolution. It also used cross-attention to accommodate conditioning information without requiring every condition to share the image’s spatial layout.\n\nStable Diffusion brought this design to a much larger user and developer community in 2022. Publicly distributed model weights supported local inference, modification, and integration into creative software. Subsequent systems changed the text encoders, network architecture, training resolution, and sampling objectives. The defining property remained the location of generative modeling in a learned representation: latent diffusion is a family of methods, not a single checkpoint, software package, or denoising architecture."
    },
    {
      "heading": "Perceptual compression and the first-stage model",
      "text": "The compression stage typically uses an encoder–decoder architecture. For an input image \\(x\\), an encoder \\(E\\) produces a latent representation \\(z=E(x)\\), and a decoder \\(D\\) produces a reconstruction \\(\\tilde{x}=D(z)\\). The latent is usually a spatial tensor rather than a single vector. Its grid preserves approximate correspondence with image regions, allowing later networks to manipulate composition and local structure.\n\nFor an image of height \\(H\\), width \\(W\\), and three color channels, a spatial downsampling factor \\(f\\) and latent channel count \\(c\\) produce a representation of approximate shape\n\n\\[z\\in\\mathbb{R}^{(H/f)\\times(W/f)\\times c}.\\]\n\nA familiar configuration in early Stable Diffusion systems maps a \\(512\\times512\\) RGB image to a \\(64\\times64\\times4\\) latent tensor. This reduces the number of scalar values from 786,432 to 16,384, a factor of 48. That ratio is not a complete measure of computational savings: network widths, attention operations, memory traffic, and decoder costs also matter. Nevertheless, it illustrates why repeated denoising can be substantially cheaper after compression.\n\nThis representation is not lossless compression. The encoder discards some information, and the decoder learns how to reconstruct plausible visual detail. Training commonly combines a pixel-level reconstruction term with a perceptual loss that compares features extracted by another neural network. An adversarial objective can encourage sharper local texture. These losses favor perceptually useful reconstructions rather than exact recovery of every original pixel. Consequently, the latent representation is shaped by the training objectives and data distribution, not by a universal definition of visual importance.\n\nRegularization controls the structure of the latent space. The original latent diffusion work considered both a weak Kullback–Leibler regularization, related to variational autoencoders, and a vector-quantization-based alternative. These are distinct choices: a latent diffusion model need not use a discrete codebook, and the widely used continuous-latent implementations should not be described as inherently quantized. Implementations also commonly rescale latent values so that their variance is appropriate for the subsequent noise schedule.\n\nCompression establishes a fundamental trade-off. Stronger downsampling reduces generation cost but can remove small text, thin lines, subtle facial detail, and other high-frequency information. Weaker compression preserves more information while increasing the cost of denoising. The decoder also imposes a reconstruction constraint: even an excellent latent generator cannot reliably recover information that the representation systematically fails to retain. For this reason, the first-stage model is evaluated separately from the diffusion prior, and improving it can materially improve the entire system without changing the denoiser."
    },
    {
      "heading": "Forward corruption and learned reverse dynamics",
      "text": "The diffusion stage learns to generate latent representations resembling those produced by the encoder. Let \\(z_0\\) denote a clean training latent. A conventional discrete forward process adds Gaussian noise according to a sequence of variances \\(\\beta_t\\):\n\n\\[q(z_t\\mid z_{t-1}) = \\mathcal{N}\\!\\left( \\sqrt{1-\\beta_t}\\,z_{t-1}, \\beta_t I \\right).\\]\n\nWriting \\(\\alpha_t=1-\\beta_t\\) and \\(\\bar{\\alpha}_t=\\prod_{s=1}^{t}\\alpha_s\\), a noisy latent at any selected timestep can be sampled directly:\n\n\\[z_t = \\sqrt{\\bar{\\alpha}_t}\\,z_0 + \\sqrt{1-\\bar{\\alpha}_t}\\,\\epsilon, \\qquad \\epsilon\\sim\\mathcal{N}(0,I).\\]\n\nThis closed form is important during training because it avoids simulating every preceding corruption step.\n\nA neural network receives the noisy latent, a representation of the noise level or timestep, and optionally a condition \\(c\\). In the common noise-prediction formulation, it estimates \\(\\epsilon_\\theta(z_t,t,c)\\). A simplified training objective is\n\n\\[\\mathcal{L}_{\\mathrm{noise}} = \\mathbb{E}_{z_0,t,\\epsilon,c} \\left[ \\left\\| \\epsilon-\\epsilon_\\theta(z_t,t,c) \\right\\|_2^2 \\right].\\]\n\nThe condition might be a text representation, a class label, or an encoded image. Training therefore teaches the model how noise differs from meaningful structure at many corruption levels, rather than teaching a fixed sequence of image-editing operations.\n\nAt generation time, the process begins with a sample approximately drawn from a standard Gaussian distribution. A sampler repeatedly uses the network’s predictions to construct a less noisy latent. After the final step, the decoder maps the resulting latent to an observable image. The reverse process is learned statistically: it does not recover an unknown original image from arbitrary noise. Instead, it produces a new sample from the distribution represented by the model, optionally biased toward the supplied condition.\n\nNoise prediction is only one parameterization. A network can instead predict the clean latent, a velocity-like combination of signal and noise, or another quantity from which the sampler derives its update. For a common convention with \\(z_t=\\alpha(t)z_0+\\sigma(t)\\epsilon\\), velocity is defined as \\(v_t=\\alpha(t)\\epsilon-\\sigma(t)z_0\\). The choice interacts with signal-to-noise ratios, loss weighting, and numerical stability. An implementation must match its prediction target to its sampler; treating an output trained as velocity as though it were noise generally gives incorrect updates.\n\nContinuous-time descriptions express related processes through stochastic differential equations and associated probability-flow ordinary differential equations. These connect diffusion sampling to numerical integration and explain why several sampling algorithms can use the same trained network. The learned object is a noise-dependent field over latent space; the sampling algorithm determines how that field is followed to obtain a final representation."
    },
    {
      "heading": "Denoising architectures and conditioning mechanisms",
      "text": "Early influential latent diffusion systems used a U-Net, a multiscale convolutional neural network originally associated with image segmentation. Its contracting path builds features at progressively lower spatial resolutions, while its expanding path restores resolution. Skip connections transfer features between corresponding levels. In a diffusion model, the input and output generally have the same latent-grid dimensions, and the network estimates a denoising-related quantity at every location.\n\nTimestep information is embedded into a vector and injected into network blocks, often through additive biases or modulation of normalized activations. This makes a single network behave differently at different noise levels. At high noise, little local evidence survives and broad structure must be inferred largely from learned statistical regularities and conditions. At low noise, the network can make more localized corrections. This distinction is a tendency rather than a strict separation into “composition” and “detail” phases.\n\nFor text conditioning, a text encoder converts a prompt into a sequence of embeddings. Early Stable Diffusion versions used text encoders associated with CLIP, while other systems use different or multiple encoders. Tokenization and the encoder’s sequence limits affect which information reaches the denoiser. The prompt is not generally supplied as raw characters to the image network, and a text embedding does not encode an unambiguous scene specification.\n\nCross-attention connects image features to these text representations. If queries \\(Q\\) are derived from visual features and keys \\(K\\) and values \\(V\\) from conditioning features, a standard attention operation is\n\n\\[\\operatorname{Attention}(Q,K,V) = \\operatorname{softmax} \\left( \\frac{QK^\\top}{\\sqrt{d_k}} \\right)V.\\]\n\nDifferent spatial positions can therefore combine different parts of the conditioning sequence. This is more flexible than appending a single text vector to every image position, although it does not guarantee correct object counting, attribute binding, or spatial relations.\n\nConditioning need not be linguistic. Class labels can be embedded directly; masks and low-resolution images can be concatenated with noisy latents; and separate branches can process depth, pose, edges, or segmentation maps. ControlNet, introduced in 2023, exemplified the use of an additional trainable branch to supply spatial control while retaining a pretrained diffusion backbone. Such mechanisms distinguish semantic instructions from geometric constraints: “a person raising an arm” and a precise pose map contain different kinds of information.\n\nLater latent generators increasingly adopted the transformer architecture. Diffusion transformers process patches or tokens derived from latents and use attention-based blocks rather than relying primarily on a U-Net’s convolutional hierarchy. Some architectures jointly process text and image token streams. The compression stage and the generative backbone are separable design choices: replacing a U-Net with a transformer does not, by itself, change whether generation occurs in latent space."
    },
    {
      "heading": "Training data, optimization, and model adaptation",
      "text": "Training normally proceeds in stages. First, an autoencoder learns to reconstruct images. Its parameters are then held fixed while the generative model learns the distribution of encoded examples. Text encoders may also remain frozen, although this is not mandatory. Freezing components limits the number of trainable parameters and stabilizes the meaning of the latent representation throughout diffusion training. Joint or partially joint training is possible, but changes the optimization problem and can make the representation itself a moving target.\n\nThe training data determine which visual concepts and image–text relationships the system can learn. Text-to-image systems commonly use paired images and captions assembled from large collections. LAION datasets, for example, provided large-scale image–text metadata used in the development of influential open-weight systems. A web caption may describe only one aspect of an image, contain surrounding page text, or omit important objects. Consequently, the diffusion model learns from an imperfect relationship between linguistic descriptions and visible content.\n\nData preparation typically includes resolution filtering, duplicate detection, cropping, and measures intended to remove unsuitable material. Caption quality and sampling balance can matter as much as raw collection size. Cropping an image without adjusting its caption may remove the object named in the text. Repeated images can receive disproportionate influence. A model trained mainly on certain visual traditions or demographic patterns may reproduce those imbalances, even when its prompts do not explicitly request them.\n\nEach optimization step samples examples, encodes them, selects noise levels, adds noise, predicts a target, and updates parameters through backpropagation. The Adam optimizer or variants such as AdamW are commonly used. Mixed-precision arithmetic and distributed computing make large runs practical across multiple accelerators. Cached latents and text embeddings can avoid repeated encoding, although caching may restrict online augmentation or prevent training the cached encoders. Noise-level sampling, target parameterization, and loss weighting determine which denoising regimes receive the strongest learning signal.\n\nConditional models often drop the condition for a fraction of training examples. The same network thereby learns both conditional and unconditional predictions, enabling guidance at inference time. This procedure should be distinguished from ordinary omission caused by missing captions: deliberate conditioning dropout is part of the model’s training design.\n\nAfter pretraining, fine-tuning can adapt a model to a domain, style, or subject. Low-rank adaptation updates selected weight matrices through small low-rank factors instead of modifying every parameter. DreamBooth, introduced in 2022, popularized subject-driven personalization through fine-tuning with a small set of example images. Textual inversion instead learns an embedding associated with a new token while leaving the main model fixed. These approaches modify different components and have different storage costs, expressive power, and risks of overfitting to the adaptation examples."
    },
    {
      "heading": "Sampling, guidance, and computational efficiency",
      "text": "Training and sampling use the same denoising network but impose different computational patterns. Training often evaluates one randomly chosen noise level for each example. Generation requires a sequence of evaluations, making the number of network calls a major determinant of latency. The original discrete diffusion formulation could involve hundreds or thousands of reverse steps; later samplers obtained useful results with substantially fewer evaluations.\n\nDenoising diffusion implicit models, introduced by Jiaming Song, Chenlin Meng, and Stefano Ermon in 2020, showed that models trained with a diffusion objective could support alternative sampling trajectories, including deterministic ones. Later numerical methods, such as DPM-Solver and related solvers, exploited mathematical structure in diffusion equations to improve the quality–cost trade-off. A sampler’s performance depends on its timestep schedule, the model’s prediction parameterization, guidance strength, and the distribution on which it is evaluated. There is no universally optimal step count.\n\nA widely used technique is classifier-free guidance. Let \\(\\epsilon_\\theta(z_t,t,c)\\) be a conditional prediction and \\(\\epsilon_\\theta(z_t,t,\\varnothing)\\) a prediction with the condition omitted. One common convention forms\n\n\\[\\hat{\\epsilon} = \\epsilon_\\theta(z_t,t,\\varnothing) + s\\left[ \\epsilon_\\theta(z_t,t,c) - \\epsilon_\\theta(z_t,t,\\varnothing) \\right].\\]\n\nHere \\(s=1\\) gives the ordinary conditional prediction under this convention. Larger values extrapolate away from the unconditional prediction and often increase prompt adherence, but may reduce diversity or cause exaggerated contrast, saturation, and other artifacts. Software packages sometimes use different guidance conventions, so numerical settings are not necessarily directly comparable.\n\nGuidance usually requires both conditional and unconditional predictions. These can be evaluated separately or combined into a larger batch, trading memory use against execution efficiency. Negative prompts commonly replace the empty conditioning input with a specified negative text condition. Their effect is therefore an alteration of the guidance direction, not a formal prohibition that guarantees the absence of a concept.\n\nFurther acceleration can come from knowledge distillation, which trains a student model to reproduce aspects of a more expensive teacher’s behavior. Consistency-based and adversarially assisted methods can support one-step or few-step generation. Such acceleration is not merely a change of sampler: an aggressively distilled model has undergone additional training and may have a different quality profile or guidance behavior.\n\nActual runtime also depends on hardware and implementation. Graphics processing units, efficient attention kernels, reduced numerical precision, model quantization, and memory-management strategies affect throughput and memory use. The autoencoder decoder and text encoder add costs outside the denoising loop. A random seed improves reproducibility within a controlled environment, but identical seeds do not guarantee bitwise identical outputs across different devices, numerical libraries, or software versions."
    },
    {
      "heading": "Image editing, controlled synthesis, and personalization",
      "text": "Text-to-image generation is only one use of latent diffusion. In image-to-image generation, an existing image is encoded and corrupted to an intermediate noise level before denoising under a new condition. Lower corruption generally preserves more of the input structure; higher corruption allows greater changes. The resulting relationship is statistical rather than exact: a small change in a prompt or noise realization can alter local details even when the broad composition remains similar.\n\nInpainting modifies a selected region while retaining surrounding content as context. An implementation may provide the denoiser with a mask and an encoded masked image, or repeatedly blend appropriate noised versions of known content into the sampling process. Dedicated inpainting models are trained with inputs that explicitly represent missing regions. Because the encoder has a finite receptive field and the latent grid is spatially compressed, a sharp pixel-space mask does not necessarily correspond to an equally sharp division in latent space.\n\nOutpainting extends an image beyond its original boundary. The challenge is not simply to extrapolate texture: the model must maintain perspective, lighting, object continuity, and scene semantics in newly created areas. Likewise, super-resolution conditions synthesis on a lower-resolution image, but recovered high-frequency details are estimates drawn from learned patterns. A generated brick texture or facial feature may look plausible without corresponding to anything actually present in the source. This distinction is especially important when images are treated as documentary evidence.\n\nSpatial control systems add information such as edge maps, depth estimates, skeletal poses, or semantic labels. These constraints can make composition more predictable than text alone. However, they inherit errors from any upstream estimator: incorrect depth or pose information can push the generator toward geometrically inconsistent results. Multiple conditions may also conflict, requiring the network or sampling procedure to balance incompatible signals.\n\nPersonalization seeks consistent depiction of a particular subject, product, or style across new scenes. Fine-tuning and learned embeddings offer one route; image-reference encoders and attention-based adapters offer another. Consistency is difficult because identity is distributed across many visual features, while poses, illumination, and backgrounds should remain variable. Excessive adaptation can reproduce training-image compositions rather than learn a reusable subject representation.\n\nThese capabilities have been incorporated into image-processing and computer graphics workflows for concept design, texture creation, compositing, and iterative visual exploration. They do not imply the precise object-level editability of a structured graphics scene. A latent image representation generally lacks explicit, persistent records for every object, material, light source, and geometric relationship, so changing one feature can unintentionally affect others."
    },
    {
      "heading": "Evaluation, failure modes, and responsible deployment",
      "text": "Latent diffusion systems must be evaluated along several distinct dimensions: reconstruction quality, sample realism, diversity, condition adherence, computational cost, and behavior on sensitive or unfamiliar inputs. An autoencoder can reconstruct well while the diffusion prior generates poorly, or the generator can produce attractive samples while failing to represent the full data distribution. Separating these components helps identify whether a defect originates in compression, denoising, conditioning, or decoding.\n\nFréchet inception distance compares the means and covariances of feature representations for generated and reference images. It is widely used for distribution-level evaluation but does not directly measure whether an individual image obeys its prompt. Its value depends on the feature extractor, reference set, sample count, and preprocessing. Text–image similarity metrics based on CLIP provide another signal, but can miss incorrect counts, swapped attributes, and implausible spatial arrangements. Human evaluation captures additional qualities, although its results depend on the task definition and rating procedure.\n\nCommon failures include malformed anatomy, repeated structures, incorrect reflections, unreadable text, and inconsistent geometry. Progress in model scale and training has reduced some of these problems without eliminating the underlying difficulty of enforcing global constraints. A denoising objective rewards statistical prediction across noisy examples; it does not explicitly enforce a scene graph, physical simulation, or symbolic specification. Strong perceptual realism can therefore coexist with structural errors that become apparent only on closer inspection.\n\nGeneralization also varies with the input. Unusual combinations of familiar concepts may fail even when each concept is individually well represented. Different languages, cultural references, technical diagrams, and rare visual styles can receive uneven treatment. Evaluation on a narrow test set may conceal such weaknesses, particularly when test prompts resemble highly frequent training descriptions. Duplicate or near-duplicate material across training and evaluation collections can further distort apparent performance.\n\nMemorization is distinct from ordinary stylistic similarity. Research has demonstrated that some diffusion models can reproduce recognizable training examples under particular conditions, especially where examples are duplicated or strongly associated with specific text. This creates concerns involving data privacy and intellectual property, but an individual output’s provenance cannot generally be inferred from visual resemblance alone. Latent compression is not an anonymization mechanism: information about training examples can remain in the complete system.\n\nDeployment controls may include training-data filtering, prompt moderation, output classifiers, access restrictions, and provenance mechanisms. These address different stages and have different limitations. A content classifier can miss prohibited outputs or reject permitted ones; a watermark can be weakened by transformations; and a cryptographic provenance record establishes a documented chain only when that chain is maintained. Photorealistic synthesis also supports deepfake creation, making consent, impersonation, and the distinction between generated and recorded imagery important operational concerns."
    },
    {
      "heading": "Extensions to video, audio, and alternative generative dynamics",
      "text": "The central idea of latent diffusion is not specific to still images. It applies whenever a learned representation substantially reduces the cost of generative modeling while preserving the information needed for reconstruction. Extending it to another domain requires an appropriate encoder–decoder pair, a denoising architecture suited to the representation, and conditioning that reflects the intended task. The quality of a representation for classification does not automatically make it suitable for high-fidelity generation.\n\nVideo adds a temporal dimension. A system can encode frames independently, encode short clips with a spatiotemporal autoencoder, or combine spatial compression with temporal compression. The denoiser must then model motion and appearance across time, often through temporal convolutions or attention. Independently generating plausible frames is insufficient: objects must retain identity, occlusions must evolve coherently, and motion should remain consistent with the scene.\n\nVideo generation also magnifies computational cost. A latent sequence contains many more tokens than a single latent image, and unrestricted attention over all space–time positions can be expensive. Factorized attention, restricted temporal windows, hierarchical generation, and stronger compression address different parts of this problem. Each introduces trade-offs. Temporal compression may lose brief events, while chunked generation can create discontinuities at boundaries. Models such as Stable Video Diffusion, introduced in 2023, illustrated how pretrained image representations and generative models could be extended to video through additional training and temporal modeling.\n\nAudio systems use representations adapted to sound rather than images. AudioLDM, introduced in 2023, applied latent diffusion to audio generation using learned audio representations and text-related conditioning. Other systems diffuse over spectrogram latents or representations produced by neural audio codecs. A decoder or vocoder converts the generated representation into a waveform. Temporal resolution, pitch structure, stereo relationships, and phase-sensitive detail create constraints that differ from those of visual synthesis. A short sound effect and a long musical passage also require different forms of temporal organization.\n\nAlternative generative objectives complicate terminology. Flow matching trains a vector field that transports samples between distributions, and rectified-flow methods seek trajectories that can be followed efficiently. These approaches are mathematically related to continuous-time generative modeling but are not identical to conventional noise-prediction diffusion training. They can operate in the same kinds of compressed image or video representations and use similar transformer backbones. Consequently, a system may belong to the broader lineage of latent generative models without being a denoising diffusion probabilistic model in the strict sense.\n\nAs these architectures develop, several design choices remain independent: the representation may be continuous or discrete; the backbone may use convolutions or transformers; the objective may predict noise, velocity, or a transport field; and inference may require many steps or a distilled few. Precise descriptions therefore identify both the learned representation and the generative dynamics rather than treating “latent diffusion” as a synonym for every modern image generator."
    }
  ],
  "links": [
    "generative-artificial-intelligence",
    "latent-space",
    "diffusion-model",
    "autoencoder",
    "variational-autoencoder",
    "generative-adversarial-network",
    "cross-attention",
    "encoder-decoder-architecture",
    "lossless-data-compression",
    "convolutional-neural-network",
    "clip-model",
    "tokenization",
    "transformer-architecture",
    "training-data",
    "backpropagation",
    "adam-optimizer",
    "distributed-computing",
    "fine-tuning",
    "low-rank-adaptation",
    "overfitting",
    "denoising-diffusion-implicit-model",
    "classifier-free-guidance",
    "knowledge-distillation",
    "graphics-processing-unit",
    "image-processing",
    "computer-graphics",
    "test-set",
    "data-privacy",
    "intellectual-property"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}