Generative AI

Synthesis samples an instance rather than recovers a specific observed person. A text-to-image model creates a scene compatible with a description; a medical volume generator creates a possible anatomy under supplied conditions. The input can be noise alone or a combination of noise and metadata, while the output may be pixels, voxels, geometry or texture. Latent diffusion is a representative visual implementation [7]. In medicine, MAISI conditions volumetric generation on body region and spacing and provides mask-guided or tumor-filling paths [9]. Texture is another object coordinate: Text2Tex generates surface appearance on an already supplied mesh, so it does not demonstrate joint geometry recovery [8]. The elementary question is which attributes the conditions control and which variations remain free, not simply whether the rendered result looks realistic.

Synthesis is useful when its variation supports a real downstream need. Training augmentation requires improvement on held-out real cases, not performance on generated labels alone; cohort construction requires representative coverage, not sample count alone. MAISI includes external autoPET image-distribution evaluation, but that is not an independent diagnostic endpoint or privacy audit [9]. MAISI-v2 combines rectified flow, region-specific losses and changed data/training conditions, so differences cannot all be attributed to one objective [27]. Conditional anatomy and rare congenital topology also have existing implementations [28, 29]. A responsible comparison measures conditional accuracy, tail support, memorization risk and real-task usefulness at matched resources. If resampling or a statistical shape model already supplies the required variation, a larger learned generator must justify its extra assumptions and costs.

Figure 3 places 20 representative synthesis publications in the common 2018–2026* window; the arrows resolve to the bibliography [30, 31, 32, 33, 34, 35, 36, 37, 38, 7, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48]. The common selection and date conventions are defined in Section 1.

Translation changes the domain while trying to preserve the relevant object. Turning a label map into a photograph changes appearance; turning MRI into synthetic CT changes an imaging representation; virtual staining maps an unstained measurement to stain-like contrast. Paired translation learns from aligned examples, whereas unpaired translation learns from marginal domain samples with additional constraints [10, 11]. The distinction is consequential: cycle consistency can support an invertible mapping without uniquely identifying the medically correct subject-wise mapping. Virtual staining demonstrates a genuine medical translation task, with registration and tissue preparation affecting the reference pairs [13]. The output should therefore be described as a conditional estimate under a specific acquisition and training protocol, not a new measurement of every biological quantity that the target modality might reveal.

The preservation contract determines the evaluation. For synthetic CT, density and geometric errors matter through dose calculation and positioning, while for staining they matter through structures and diagnostic observations. AortaDiff already couples contrast translation with segmentation and clinically interpretable measurements, so “task-aware joint translation” is not an untouched idea [49]. Its preprint evaluation also has ROI, alignment and derived-reference limitations that should remain visible. The SynthRAD2025 report separates image, geometric and dose criteria and demonstrates why rankings depend on the endpoint [50]. Unpaired realism or average similarity is insufficient when a rare structure changes treatment planning. Translation should be compared against the current clinical workflow, including non-generative alternatives, and failures should trigger an explicit fallback rather than be hidden in a favorable mean score.

Figure 4 places 20 representative translation publications in the common 2018–2026* window; the arrows resolve to the bibliography [51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70]. The common selection and date conventions are defined in Section 1.

Editing asks for a controlled difference from an existing instance. An instruction such as changing an object's color is not unconditional sampling; the original image supplies identity and non-target attributes. InstructPix2Pix learns from generated instruction-edit triplets and reports difficulties with unwanted changes, isolation and viewpoints [14]. In medicine, digital siblings already create local anatomic variations and assess morphology and topology [15]. Here, the useful output is often a family of controlled alternatives for sensitivity analysis rather than a clinically asserted future patient. One must specify the editable region, required invariants, permitted topology changes and the meanings of edit magnitude. A visually small change can still be large for a downstream measurement, especially near a lesion, valve or vessel boundary.

A counterfactual image is not automatically a causal treatment counterfactual. An edited anatomy can support a computational experiment if the simulator and intervention are explicitly defined, but visual plausibility alone does not identify how a patient would respond to treatment. Distributional editing also cannot establish that all non-target physiology is preserved. Appropriate tests compare independent measurements before and after editing, stress boundary cases, and quantify collateral changes outside the requested region. Classical parameterized deformation or resampling can be a strong comparator when the desired variation is simple. The contribution of a learned editor should be a newly supported, validated class of variation or a better error–cost trade-off, not the relabeling of existing anatomy manipulation. This distinction keeps controlled sensitivity studies useful without overstating biological causality.

Figure 5 places 20 representative editing publications in the common 2018–2026* window; the arrows resolve to the bibliography [71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 14, 81, 82, 83, 84, 85, 86, 87, 88, 48]. The common selection and date conventions are defined in Section 1.

Restoration retains an observed scene while compensating for degradation. Denoising, deblurring, colorization and filling missing regions differ in how much of the answer is identifiable. Palette uses a shared image-conditioned diffusion formulation across several such tasks [16]; multiple conditions do not alone establish multitask learning without shared task training and evaluation. DDRM reuses an image prior across linear inverse problems through operator structure [17]. Medical examples include microscopy restoration and MAISI tumor filling [18, 9]. A filled region may be a plausible simulation object rather than evidence of the true hidden anatomy. The task description must say whether the aim is retrospective recovery, prospective diagnosis, synthetic augmentation or experimental perturbation, because identical-looking outputs can have very different acceptable failure modes.

The unobserved region is where a strong prior can become a liability. A small residual on visible pixels says little about missing structures; several plausible samples do not automatically quantify calibrated uncertainty. Existing work already separates prior-induced hallucinations from ordinary reconstruction error and develops distribution-referenced hallucination metrics [89, 18]. The latter depends on a defined reference and approximation assumptions, not a universal truth oracle. Evaluation should therefore distinguish measurement-supported detail, ambiguous completion and wrong but convincing structures. For a clinical use, the question is whether the ambiguity changes a diagnosis or quantitative endpoint and whether the system can decline to fill it. Deterministic restoration remains an essential comparator: this operation-based taxonomy includes such comparators without claiming that every denoiser is a stochastic generative model.

Figure 6 places 20 representative restoration / fill publications in the common 2018–2026* window; the arrows resolve to the bibliography [90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 16, 100, 101, 102, 103, 104, 105, 106, 107, 108]. The common selection and date conventions are defined in Section 1.

Reconstruction uses a specified forward model and a particular measurement. CT projection data and MRI k-space constrain a latent object through Equation 2; they are not merely prompts for a similar-looking image. Generative compressed sensing makes representation error explicit, while DPS approximates a noisy likelihood using denoised estimates [5, 19]. Medical score-based reconstruction demonstrates reuse of an unconditional prior under different sampling operators [20]. In the inspected study, CT projections and single-coil MRI measurements are simulated from image data. Those experiments probe operator variation, but they do not constitute prospective scanner or diagnostic validation. A precise report separates real source images, simulated measurements, real acquisition tests, external patients and clinical readers rather than collapsing all of them into “real medical data.”

Faithfulness requires more than exact data fitting. Noise and model mismatch can make strict fitting undesirable, while a small data residual cannot certify components poorly constrained by the operator [89]. Generator-range bias can persist even with more observations [5], and posterior-sample diversity can mix model, numerical and observation effects. A stronger experiment fixes the raw-data split, compares a competitive non-generative method and a generative baseline, and measures a prespecified lesion, quantitative or reader endpoint at a realistic compute budget. Active acquisition is already explored with diffusion posterior samples [109]; novelty must lie beyond merely choosing the next view. Uncertainty becomes practically useful only when it supports a calibrated decision, rejection or additional measurement with a favorable total burden, including failed cases and reacquisition time.

Figure 7 places 20 representative measurement inverse publications in the common 2018–2026* window; the arrows resolve to the bibliography [110, 111, 112, 113, 114, 115, 20, 17, 116, 19, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126]. The common selection and date conventions are defined in Section 1.

Time introduces several distinct problems. Unconditional video creates a sequence; interpolation fills missing times; prediction conditions on a past; action-conditioned generation also receives an intervention on the acquisition or environment. Video diffusion and Genie exemplify joint visual sequences and interactive dynamics [21, 22]; 4D-fy adds a dynamic 3D representation with complementary image and video supervision [23]. Medical cases include CHeart's within-cycle anatomy, Cardiac Mesh Flow's dynamic meshes and ultrasound views under probe movement [24, 25, 26]. A cardiac cycle is not longitudinal disease progression, and an autoregressive “causal” architecture is not a treatment-effect model. Similarly, a smooth rendered mesh is not automatically a measured displacement field. Each temporal use must specify its conditioning history, horizon, action semantics and independent reference.

Temporal validity needs negative controls that defeat superficial metrics. Cardiac Mesh Flow's ablation shows that static hearts can attain strong periodic consistency, making endpoint agreement an insufficient motion validator [25]. CHeart's dense temporal anatomy relies on processed segmentation labels, not independent local motion truth [24]. SonoWorld already grounds ultrasound generation in metric poses, while EchoWorld studies predictive representations for probe guidance [26, 127]; neither inspected evaluation establishes an independent real-robot clinical policy-transfer trial. Contact force, deformation, physiology and scanner settings can all break a pose-only observation model. Tests should include static outputs, condition shuffling, non-replayed actions and local motion or control endpoints. The important advance is not longer or smoother video, but a demonstrably correct relation between time, action and the downstream quantity.

Figure 8 places 20 representative temporal generation publications in the common 2018–2026* window; the arrows resolve to the bibliography [128, 129, 130, 131, 21, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146]. The common selection and date conventions are defined in Section 1.

Conditional. Conditional generation asks how outputs change when information is supplied, not how much training data are available. In ML, a condition selects a distribution: the same class or context can still admit many valid samples. SPADE transfers this idea to CV by synthesizing scenes from semantic layouts. MAISI provides a medical counterpart, generating synthetic three-dimensional CT with organ segmentations as conditions and configurable voxel spacing. The desired invariant is agreement with the supplied anatomy; unsupplied appearance remains variable. Text conditioning is one special case, while controllability additionally asks whether specific changes can be steered reliably. Suggested checks are condition agreement, diversity and anatomical plausibility, not assumed guarantees. A conditioning input does not establish patient identity, pathology truth, privacy or clinical benefit. The timeline and examples draw on the following sources [52, 30, 150, 151, 147, 152, 37, 97, 16, 7, 39, 153, 41, 148, 154, 149, 155, 9, 156, 157].

Text-conditioned. Text-conditioned generation uses language as the condition; it is therefore compatible with other conditioning modalities and data regimes. In ML, words specify requested attributes and relations, but the generator must translate them into visual content. Imagen provides a CV anchor for text-to-image diffusion, making object counts, colors and relations useful teaching probes rather than guaranteed successes. RoentGen adapts a pretrained latent diffusion model using chest radiographs and reports, then generates radiographs from medical-language prompts. Suggested checks distinguish prompt adherence from image realism and independently assess depicted findings and anatomy. A requested diagnosis is neither evidence that the output depicts it correctly nor truth about a patient. These abstract-supported examples establish generation capabilities, not complete pathology control, patient preservation or clinical proof. The timeline and examples draw on the following sources [151, 152, 37, 158, 7, 39, 153, 40, 41, 42, 148, 46, 159, 47, 160, 161, 162, 163, 156, 157].

Controllable. Controllability is a steering objective, whereas conditioning describes information supplied to a generator. A useful ML probe varies one requested factor while fixing other inputs and, when meaningful, the random seed; outputs must still be checked for coupled changes. ControlNet adds trained spatial controllers to a pretrained text-to-image diffusion model using pose, edge or depth conditions. MAISI documents organ-mask conditioning for CT synthesis, providing a narrower anatomical-control example rather than a disease-intervention model. Suggested checks quantify response to the changed mask and unintended changes elsewhere. Neither a frozen base generator nor an anatomical control implies training-free control learning, causal identification, preserved patient identity or clinical validity. These coordinates can overlap with text conditioning and with either low-data adaptation or extensive training. The timeline and examples draw on the following sources [147, 71, 72, 73, 74, 164, 153, 78, 14, 154, 165, 149, 155, 83, 166, 167, 168, 9, 169, 170].

Few-shot. Few-shot generation limits examples of a specified target, such as a subject or domain, rather than all data ever used. In ML, adaptation transfers a learned prior to that small support set; other methods can condition on support without weight updates. DreamBooth illustrates CV personalization by fine-tuning a pretrained text-to-image model on a few subject photographs. EndoFSA documents a medical case: a generator pretrained on abundant normal capsule-endoscopy images is adapted to scarce abnormal domains through restricted parameter updates. These are generated domain examples, not reconstructions of a particular patient's lesion. Suggested checks include unseen-patient lesion review, diversity and memorization auditing. Exact medical shot counts and independent clinical performance require further evidence; few trainable parameters alone would not establish few-shot learning. The timeline and examples draw on the following sources [174, 175, 176, 6, 177, 178, 171, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 170, 172].

Zero-shot / training-free. This coordinate concerns reuse without fitting generator weights to the target condition or acquisition operator. It differs from semantic zero-shot recognition of unseen categories. No target labels also does not mean no pretraining or no target-modality images. In ML, an existing prior is combined with a new condition during inference. RePaint turns a trained unconditional diffusion model into an inpainter by changing reverse-sampling iterations rather than retraining its network. In medicine, Score-MRI combines a magnitude-image-trained score prior with data-consistency steps across sampling patterns. Its training-free aspect is acquisition-specific reuse, not learning MRI anatomy from nothing. Suggested tests assess measurement residuals and independently checked anatomy. Inference projections or latent optimization still incur computation; agreement with measured data cannot verify unmeasured lesions. The timeline and examples draw on the following sources [76, 64, 115, 100, 17, 78, 19, 80, 82, 117, 165, 139, 83, 190, 166, 167, 168, 169, 191, 192].

Data-efficient. Data efficiency asks whether useful generation can be learned with fewer real examples or less target supervision at a specified quality level. In ML, regularization or transfer can reduce overfitting, but the data budget must be stated. ADA provides a CV anchor through adaptive discriminator augmentation for limited-data GAN training, including training from scratch and transfer. EndoFSA illustrates medical target-data reuse: normal-image pretraining supports adaptation to scarce abnormal endoscopy images. This is a documented low-data design, not a verified cross-method efficiency ranking. Suggested tests plot quality, diversity and independent downstream performance against real patient counts with fixed splits and comparable pretraining access. Count abundant pretraining separately; faster sampling, smaller adapters and more synthetic images do not themselves demonstrate lower real-data requirements. The timeline and examples draw on the following sources [174, 175, 176, 173, 193, 194, 195, 196, 6, 197, 177, 198, 178, 171, 180, 199, 200, 201, 172, 202].

Multitask. Multitask describes how objectives share learning, not how many channels enter a network. A generic ML example shares parameters across conditional generation objectives while retaining task-specific evaluation under shared or task-specific losses. Palette provides a CV instance: one generalist diffusion model handles colorization, inpainting, uncropping, and JPEG restoration. In medicine, AortaDiff combines synthetic contrast-enhanced CT generation from noncontrast CT with lumen and thrombus segmentation. It therefore also overlaps joint generation; coordinate values are not exclusive categories. Both image quality and segmentation or measurement error require separate checks against specialist baselines. The intended invariant is anatomical identity across tasks, not identical pixel values. Shared architecture alone does not establish a single unchanged checkpoint, and benchmark performance does not prove clinical substitution. The timeline and examples draw on the following sources [76, 16, 7, 39, 207, 208, 209, 210, 211, 203, 212, 213, 214, 215, 216, 217, 204, 218, 219, 220, 49].

Multimodal. Multimodal generation connects semantically different representations, such as language and images. In ML, generating an image from text models a cross-modal conditional distribution; the text may remain fixed while only the image is sampled. UniDiffuser illustrates the CV distinction because the same framework supports image-to-text, text-to-image, and paired image-text generation through different sampling settings. UniX supplies a medical example with chest-radiograph report generation and text-guided radiograph synthesis. The invariant to test is agreement about the requested findings, not pixelwise identity with an individual patient. Multimodal conditioning alone is not joint sampling, and several MRI contrasts do not by themselves establish an any-to-any system. Required checks include finding-level omissions, contradictions, and unsupported details; a realistic image or fluent report is not clinical proof. The timeline and examples draw on the following sources [37, 39, 41, 203, 221, 222, 223, 212, 224, 213, 225, 226, 227, 214, 228, 229, 217, 204, 218, 230].

Joint generation. Joint generation concerns dependence between outputs belonging to one sample. In ML, learning two plausible marginal distributions is insufficient: an image and its annotation must agree when sampled together. JointNet illustrates this in CV by coupling RGB and dense-modality diffusion branches to generate corresponding RGB-depth outputs. MedSegFactory provides a medical counterpart with two streams for images and segmentation masks, connected by joint cross-attention during denoising. Coupling can also use a shared latent or a dependent pipeline; simultaneous execution is not the defining requirement. Required checks assess each output and their correspondence, including boundaries, label semantics, and downstream performance on independent real cases. A jointly generated mask is not independent ground truth: both streams may reproduce the same anatomical error, and agreement alone cannot establish clinical validity. The timeline and examples draw on the following sources [231, 232, 233, 234, 203, 235, 221, 236, 205, 237, 238, 239, 240, 229, 206, 241, 242, 243, 244, 230].

Image. Image specifies the generated object, not the learning mechanism or task. In ML, a model samples an array whose values represent intensity or color; generation can be unconditional, conditioned, or part of restoration. StyleGAN supplies a CV example by synthesizing face images with controllable appearance variation. MAISI demonstrates the medical target by using latent diffusion to synthesize three-dimensional CT, with organ segmentations available as ControlNet conditions. An intensity volume remains an image even though it occupies three-dimensional space; a segmentation volume instead encodes anatomical labels or shape. Required checks include anatomical agreement, intensity bias, diversity, and memorization rather than visual realism alone. An organ mask is a control signal, not clinical evidence, and a synthetic CT must not be treated as an acquired patient measurement. The timeline and examples draw on the following sources [30, 31, 32, 33, 34, 36, 37, 38, 7, 39, 40, 41, 44, 46, 159, 47, 247, 160, 161, 219, 9].

Texture. Texture names generated appearance, which may be a spatial pattern, a mesh color map, or a material field. The ML problem models spatial appearance statistics rather than geometry alone. Text2Tex is a direct CV example: a supplied mesh and text guide multiview diffusion inpainting to build its texture. Mesh geometry is the fixed support; seams and cross-view agreement need checking. The medical bridge here is explicitly an appearance analogy: virtual staining maps label-free tissue autofluorescence to a stained-looking microscopy image using a GAN. It generates image appearance, not a UV texture for a mesh. Likewise, MIVAE couples MRI appearance with cardiac shape; that does not make MRI intensity a surface material. Matching histological structures and disease-relevant details requires independent validation beyond visual realism. The timeline and examples draw on the following sources [248, 231, 249, 250, 251, 252, 253, 234, 211, 8, 236, 254, 255, 256, 257, 258, 259, 243, 260, 261, 13, 245].

Shape. Shape specifies a geometry-bearing output, not a particular generator. In ML, learning a distribution of coordinates, occupancies, or signed distances permits new spatial configurations to be sampled. PointFlow illustrates the CV case by modeling a distribution of shapes and a conditional distribution of points for each shape. Its point clouds do not themselves provide triangle connectivity. CAN-FLOW supplies a medical example: a conditional normalizing flow models metadata-dependent cardiac shape representations and generates biventricular anatomies. It is not a flow-matching method merely because its name contains flow. Required checks cover anatomical validity, phenotype distributions, subgroup variation, and geometric representation conventions. Evidence from healthy training anatomies does not establish diseased anatomy coverage, physiological function, or simulation readiness. Appearance and temporal motion remain separate, potentially overlapping coordinates. The timeline and examples draw on the following sources [262, 246, 263, 264, 265, 266, 234, 267, 268, 269, 210, 270, 271, 272, 273, 274, 275, 276, 277, 220, 28].

4D / 3D+t. This coordinate asks whether three spatial dimensions evolve over physical time, rather than whether a sampler runs for many numerical steps. At the ML level, a model generates dependent spatial states indexed by time; independent volumes do not establish a coherent trajectory. In CV, MAV3D converts text into a dynamic 3D radiance field that can be rendered from different viewpoints. In medicine, CHeart conditions generated cardiac anatomical sequences on clinical factors. These examples illustrate an output structure, not an exclusive model family, and can also carry conditioning or mechanism labels. Validation should compare anatomy and motion against held-out reference sequences, including ventricular volume trajectories where appropriate. Smoothness or plausible population variation does not establish measured patient physiology, electrical activation, or a valid individual digital twin. The timeline and examples draw on the following sources [283, 284, 285, 280, 286, 278, 24, 287, 288, 23, 289, 290, 291, 292, 293, 294, 295, 296, 145, 297].

Frame interpolation. Frame interpolation fills times strictly between observed endpoints; extrapolation predicts beyond them. The ML problem is endpoint-conditioned synthesis, potentially with several plausible intermediate paths. FILM supplies the CV example by synthesizing intervening image frames even with large motion. The medical DDM example instead generates intermediate 3D cardiac MR volumes between diastolic and systolic observations using learned deformation. Thus the same temporal coordinate can describe ordinary image frames or spatial volumes, and it overlaps 4D generation when the latter evolve over time. Evaluation should withhold genuinely acquired intermediate phases and check anatomy, timing, and motion, rather than score endpoint reproduction alone. Interpolating a visually smooth heartbeat does not determine that patient's actual intermediate physiology or justify predicting disease progression between clinical visits. The timeline and examples draw on the following sources [298, 299, 300, 301, 302, 303, 279, 280, 304, 305, 306, 307, 308, 309, 310, 311, 312, 313, 314, 315].

Action / world model. An action-conditioned model predicts what observation or state follows a specified action and history; a policy instead chooses the action. World Models illustrates the ML formulation with a probability distribution over the next latent state conditioned on the current action and memory. GameNGen turns past frames and player actions into the next game frame. SonoWorld is a documented ultrasound preprint that grounds video generation in metric probe poses using EM-tracked freehand scans. Its pose-control evaluation is not proof that every generated response to a new action is physically valid. Validation should compare held-out action responses, opposite-direction controls, and accumulated rollout drift. Replaying recorded poses or producing convincing videos does not establish causal intervention validity, safe robotic control, or clinical benefit. The timeline and examples draw on the following sources [281, 316, 317, 318, 319, 320, 321, 322, 22, 323, 324, 325, 282, 326, 327, 328, 329, 330, 331, 26].

GAN. A GAN is defined by adversarial learning between a generator and discriminator, not by a specific network architecture or imaging task. The ML example maps random latent codes into samples while discriminator feedback supplies a learned training signal. StyleGAN applies this principle to image synthesis with scale-specific style control. MedGAN applies adversarial, feature, and style losses to PET-to-CT translation for attenuation correction rather than direct CT diagnosis. Here the source scan constrains the desired anatomical content, but those losses do not guarantee that lesions or quantitative attenuation values survive translation. Evaluation should measure distribution coverage as well as realism, then check paired anatomy, quantitative errors, and task-specific reader judgments. A discriminator score is neither diagnostic confidence nor evidence that all clinically important modes are represented. The timeline and examples draw on the following sources [337, 52, 338, 53, 92, 30, 31, 147, 32, 58, 173, 193, 339, 340, 341, 342, 198, 343, 344, 345, 332].

Diffusion. Diffusion models learn how to reverse progressive corruption, commonly through denoising or noise-dependent score estimation. The ML example starts from noise and repeatedly applies the learned reverse process. Latent diffusion performs this process in a compressed representation and decodes it into an image, with text or other inputs providing optional conditions. Medical Diffusion documents three-dimensional MRI and CT synthesis. Its abstract reports reader assessment of appearance, anatomical correctness, and slice consistency; this is narrower than diagnostic equivalence or privacy protection. Validation should separately assess anatomical structure, rare abnormalities, memorization, and performance on independent clinical tasks. Numerical sampling time is not patient time. A deterministic probability-flow ODE can also sample a score-trained diffusion model, so using an ODE alone does not establish flow-matching training. The timeline and examples draw on the following sources [346, 33, 347, 34, 36, 97, 38, 7, 39, 40, 41, 348, 349, 19, 43, 44, 149, 46, 350, 351].

Flow matching. Flow matching trains a velocity field against targets defined along chosen source-to-data probability paths. At the ML level, the input is an intermediate sample and path time, and the output is a transport direction used to generate new samples. Rectified-flow transformers provide the CV example through text-to-image synthesis. Flow Matching for Medical Image Synthesis documents an optimal-transport variant for 2D and 3D outputs with class or mask conditioning. This is a training coordinate, not a synonym for any invertible network, optical-flow estimator, or ODE solver. It can share probability paths and sampling machinery with diffusion. Validation should check transport accuracy, condition adherence, and anatomy at a declared sampling budget. The abstract-supported medical application does not establish clinical benefit, exact anatomical preservation, or universally faster generation. The timeline and examples draw on the following sources [352, 353, 335, 354, 355, 356, 357, 358, 359, 360, 47, 361, 362, 363, 364, 365, 336, 366, 367, 368].

Fidelity. Fidelity must name its reference: a data distribution, an input condition, a perceptual target, or a clinically meaningful feature. At the ML level, the density and coverage metrics associated with PRDC distinguish sample fidelity from distribution coverage rather than collapsing them into one score. GenEval supplies the CV example by checking generated object counts, colors, and spatial relationships against text prompts; it is an evaluator, not a generator. The medical human-feedback framework uses pathologist judgments, a learned reward model, and diffusion fine-tuning to improve clinical plausibility. A convincing texture can still conceal an incorrect cell type or missing abnormality. Evaluation should therefore include independent expert and task-specific checks under patient-separated testing. Expert preference or better generic image scores alone does not establish diagnostic fidelity or prospective utility. The timeline and examples draw on the following sources [372, 373, 374, 375, 369, 376, 38, 377, 378, 379, 380, 381, 382, 371, 383, 370, 384, 385, 386, 387].

Efficiency. Efficiency asks which resource is saved at what output quality, not simply how many sampling steps are removed. The ML comparison fixes data, hardware, and a quality target before reporting latency, memory, or training cost. DPM-Solver illustrates CV inference acceleration through a specialized solver for diffusion ODEs without additional training. MAISI-v2 documents accelerated 3D medical image synthesis using rectified flow and a region-specific contrastive loss; it is not merely DPM-Solver applied to medical scans. These mechanisms can occupy other coordinates at the same time. Evaluation should include encoding, decoding, loading, postprocessing, quality control, failures, and retries when measuring the total workflow. A fast sampling kernel or fewer network evaluations does not prove lower end-to-end cost, preserved lesion information, or diagnostic noninferiority. The timeline and examples draw on the following sources [347, 36, 7, 388, 389, 348, 349, 352, 390, 391, 392, 45, 393, 394, 395, 396, 397, 398, 366, 399, 27].

Measurement consistency. Measurement consistency tests a reconstruction against acquired data through a specified forward operator and noise model. The ML example constrains observed components while a generative prior supplies unresolved ones. DDNM illustrates the CV case by modifying null-space content during diffusion-based linear image restoration. The medical score-based inverse framework reconstructs CT or MRI from measurements using a physical measurement model at test time. DPS additionally illustrates why noisy observations need a noise-aware objective rather than an indiscriminate exact-fit projection. Validation should examine residuals using the stated noise statistics and compare clinically relevant structures against independent references. A small residual cannot identify anatomy in an operator's null space: different images may produce identical observations. Neither measurement fit nor a plausible posterior sample alone establishes unobserved truth or calibrated uncertainty. The timeline and examples draw on the following sources [400, 114, 115, 20, 100, 17, 116, 19, 391, 117, 401, 402, 403, 404, 121, 122, 405, 124, 406, 407].