14. AI, Neural Graphics, and Generative Content

Introduction and Context

Computer graphics has traditionally been described as the algorithmic production of images from geometric models, materials, lights, cameras, and animation data. Neural graphics extends this tradition by learning parts of the graphics process from data. Instead of specifying every rule explicitly, a neural model may learn how appearance varies with viewpoint, how incomplete observations correspond to three-dimensional structure, or how a text description can be converted into a visual asset.

This shift does not eliminate conventional graphics. Rather, it creates hybrid systems in which classical geometry, rendering, optimization, and machine learning cooperate. A production system may use a neural network to infer a scene, a differentiable renderer to optimize geometry, a conventional rasterizer to display the result, and a generative model to create missing textures. Understanding these systems therefore requires both graphics knowledge and an ability to reason about data, optimization, uncertainty, compute, controllability, and reliability.

Neural graphics is especially important in view synthesis, 3D reconstruction, virtual production, robotics, augmented and virtual reality, digital twins, medical visualization, games, and content creation. Its promise is substantial, but so are its limitations. A visually convincing result may contain incorrect geometry, unstable motion, fabricated details, or hidden dependence on biased training data. At postgraduate level, the central question is not merely how to use a model, but when its output is appropriate, how it should be evaluated, and what guarantees can reasonably be claimed.

Learning Objectives

After completing this chapter, the reader should be able to:

1. Explain the relationship between conventional, neural, differentiable, and generative graphics.

2. Describe the principles of neural radiance fields, Gaussian splatting, learned reconstruction, and diffusion-based content generation.

3. Formulate a differentiable rendering problem and identify its optimization variables and losses.

4. Select an appropriate neural graphics method for a practical application by considering quality, speed, data, compute, and controllability.

5. Critically evaluate artifacts, uncertainty, bias, reproducibility, and reliability in AI-generated visual content.

6. Design a responsible AI-assisted graphics pipeline with human oversight and suitable validation.

From Explicit Rendering to Learned Scene Functions

In a conventional renderer, an image is generated by applying a known rendering pipeline to a scene. The scene contains geometry, materials, lights, and a camera. For a surface point, the rendering equation describes outgoing radiance as a combination of emitted light and reflected incoming light. Rasterization and ray tracing implement different computational approximations to this process.

A neural graphics system can learn an unknown component of the pipeline. A network might predict depth from an image, represent radiance directly, estimate material parameters, or generate a plausible object from text. The learned component is commonly trained by minimizing a loss function over examples. If the model is represented by parameters theta and the training examples are indexed by i, a simplified objective is

L(theta) = sum_i D(F_theta(x_i), y_i) + lambda R(theta),

where F_theta is the model, D measures discrepancy between prediction and target, and R is a regularization term. Unlike a conventional algorithm, the behavior of F_theta depends on the distribution and quality of the training data.

A useful distinction is between three roles of learning. In inverse graphics, a model estimates scene variables from images. In neural rendering, a model represents or synthesizes the appearance of a scene. In generative graphics, a model produces new content that may not correspond to a particular observed scene. These categories overlap, but they imply different expectations. Reconstruction should be judged by geometric and multi-view consistency; generation should also be judged by diversity, controllability, and semantic alignment.

Important terminology

A neural scene representation is a learned function or set of learned parameters describing geometry, appearance, or both. An implicit representation defines a field queried at arbitrary coordinates rather than storing an explicit mesh or voxel grid. A latent representation compresses content into a lower-dimensional feature space. View synthesis means generating an image from a viewpoint not directly observed. Inverse rendering estimates scene properties from images. Learned reconstruction uses data-driven priors to infer missing or ambiguous structure.

These distinctions matter because a representation that is excellent for rendering may be unsuitable for editing. A radiance field can reproduce observed views without exposing clean surfaces, while a mesh supports conventional rigging and collision detection but may fail to capture complex appearance. Representation choice is therefore an engineering and scientific decision, not a minor implementation detail.

Differentiable Rendering as an Optimization Framework

A renderer is differentiable when useful derivatives of its output with respect to selected scene parameters can be computed. Let I = R(phi) denote an image produced by renderer R, where phi may include camera pose, vertex positions, texture values, light parameters, or neural-network weights. Given a target image I*, optimization seeks

min_phi L(R(phi), I*).

Gradients can be obtained analytically, through automatic differentiation, or by numerical approximation. The result allows image-space evidence to update scene parameters. For example, silhouette mismatch can move mesh vertices, photometric error can optimize camera pose, and a semantic loss can guide a generated asset toward a requested category.

Differentiability is not equivalent to correctness. Visibility changes, occlusion, shadows, discontinuities, and sampling can produce unstable or biased gradients. Soft rasterizers approximate hard visibility with smooth functions, making optimization possible but potentially changing the problem being solved. A system may converge to a low image loss while producing implausible geometry, because many different scenes can explain the same image.

The problem is frequently ill-posed. A single photograph does not uniquely determine hidden surfaces, lighting, reflectance, and scale. Effective systems therefore combine multiple views, geometric priors, regularization, masks, depth measurements, semantic constraints, or domain-specific knowledge. A practical loss may contain photometric, silhouette, depth, normal, smoothness, perceptual, and semantic terms. The weights among these terms express design priorities and can strongly influence the result.

Differentiable rendering is used for camera calibration, 3D pose estimation, inverse lighting, material capture, mesh fitting, articulated-body tracking, and neural scene optimization. It is particularly useful when the desired output is a structured scene rather than only an image. Its main advantage is a bridge between observable image evidence and editable graphics parameters.

Neural Radiance Fields and Continuous View Synthesis

A neural radiance field, commonly called a NeRF, represents a scene as a function queried at a three-dimensional position and viewing direction. The function predicts a density or opacity value and emitted color. A camera ray is sampled at a sequence of points, and the predicted values are accumulated using volume rendering. In simplified form, the pixel color is

C(r) = sum_i T_i alpha_i c_i,

where c_i is the predicted color at sample i, alpha_i is opacity, and T_i is the transmittance from the camera to that sample.

During training, camera poses and images are supplied, and the field parameters are optimized so that rendered views match the training images. Once trained, the field can synthesize novel views. The representation is continuous and can model detailed appearance without first constructing an explicit polygonal surface.

NeRF-like methods have generated extensive research on faster training, improved pose estimation, unbounded scenes, dynamic objects, surface extraction, semantic labeling, and generalization across scenes. Their original form, however, has significant limitations. Training can be computationally expensive, rendering requires many network evaluations per ray, and quality depends on accurate camera calibration and sufficient viewpoint coverage. Specular reflections, transparency, transient objects, and moving content violate assumptions of a static, view-consistent scene.

A radiance field may also memorize observations rather than recover physically meaningful structure. This is acceptable for a visual walkthrough but problematic for measurement, simulation, or editing. It is therefore essential to distinguish novel-view quality from geometric validity.

Gaussian Splatting and Explicitly Renderable Fields

Gaussian splatting represents a scene using many three-dimensional Gaussian primitives. Each primitive has a position, scale, orientation, opacity, and appearance parameters, often including view-dependent color coefficients. During rendering, the Gaussians are projected onto the image plane and blended according to depth and opacity.

Compared with a conventional coordinate-based radiance field, the primitives provide an explicit, spatially organized representation that can be rendered efficiently with specialized rasterization or splatting procedures. Training and interactive display can be substantially faster in many settings. The representation is also easier to inspect and manipulate than a deeply implicit function, although it is not automatically equivalent to a clean surface mesh.

The method introduces its own concerns. Excessive or poorly placed Gaussians may produce floaters, halos, holes, or unstable appearance outside the training viewpoints. Large scenes require memory management and level-of-detail strategies. Editing individual semantic objects can be difficult unless the representation is augmented with segmentation, object identities, or a conversion to another representation. As with radiance fields, high image quality is not proof of complete geometry or physical correctness.

The comparison between radiance fields and Gaussian splatting illustrates a broader principle: a representation should be selected according to the intended operation. If the main requirement is photorealistic view interpolation, a learned field or splat representation may be effective. If the requirement is manufacturing, collision detection, deformation, or physically based relighting, explicit geometry and materials may still be necessary.

Learned Reconstruction and 3D Generative Models

Learned reconstruction estimates three-dimensional content from incomplete observations. Inputs may include one image, multiple calibrated images, video, depth, point clouds, sketches, or language. The model uses visual priors learned from data to infer hidden or ambiguous information. Single-image reconstruction is inherently uncertain: a model may infer a plausible back side of an object, but the result is not directly supported by the image.

Modern systems use convolutional networks, transformers, implicit functions, signed distance fields, point-based representations, diffusion models, and hybrid architectures. Some predict depth and occupancy; others generate a latent 3D representation that can be rendered from multiple views. Multi-view consistency is a major design goal, because independently generated views can exhibit changing geometry, texture, or object identity.

Generative models learn a distribution of content rather than one deterministic mapping. Autoregressive models generate sequences of tokens or features, whereas diffusion models learn to reverse a controlled noise process. In visual generation, conditioning may include text, sketches, depth, edge maps, segmentation, pose, or reference images. Conditioning improves control but does not guarantee compliance. A prompt such as “a red chair viewed from the side” may be semantically interpreted while violating exact dimensions, topology, or material requirements.

For graphics production, generative models are valuable for ideation, texture and material creation, background synthesis, image completion, asset variation, and rapid concept exploration. They are less reliable when the output must satisfy exact topology, legal provenance requirements, engineering tolerances, or persistent identity across many frames. A generated image is not automatically a production-ready asset.

AI-Assisted Content Pipelines

A robust pipeline separates generation from verification and conversion. A typical workflow begins with a specification: intended use, target platform, visual style, polygon or texture budget, licensing constraints, and required physical or semantic properties. Reference material is then collected and assessed for rights, diversity, quality, and suitability.

The generation stage may create concepts, meshes, textures, animations, or scene reconstructions. Selection and refinement follow. A technical-art stage converts outputs into required formats, retopologizes geometry, generates UVs and level-of-detail versions, bakes or compresses textures, and checks coordinate conventions. Validation then tests image quality, temporal consistency, topology, performance, collision behavior, and compliance with the original specification.

Human review should not be treated as a final cosmetic step. Artists and engineers contribute domain knowledge that current models do not reliably possess. They identify impossible structures, incorrect cultural details, inappropriate representations, and hidden artifacts. Versioned prompts, model identifiers, seeds, input references, edits, and evaluation results should be recorded to support reproducibility and accountability.

In an organizational setting, AI assistance changes roles rather than simply removing them. Artists increasingly act as visual directors, reviewers, editors, and custodians of style and quality. Technical directors design constraints, asset schemas, validation tools, and integration processes. Managers must evaluate not only time saved but also review effort, error costs, licensing risk, and dependence on external services.

A Framework for Choosing a Neural Graphics Method

Method selection can be organized around five questions.

First, what is the required output? An image, a video, a navigable scene, an editable mesh, a material estimate, or a physically meaningful model each demands different representations. Second, what observations are available? A single unconstrained image is fundamentally different from calibrated multi-view imagery or a dense scan. Third, what must remain controllable? Camera motion, object identity, topology, dimensions, lighting, and animation may require explicit parameters.

Fourth, what are the operational constraints? Training time, inference latency, GPU memory, storage, bandwidth, and deployment hardware affect feasibility. Finally, what level of error is acceptable? An advertising concept may tolerate hallucinated geometry; an architectural visualization, medical interface, or robotic simulator may not.

The following general choices are useful, though not absolute:

• Use differentiable conventional rendering when structured parameters and measurable image evidence are available.
• Use radiance fields for high-quality scene-specific view synthesis when training and calibration resources exist.
• Use Gaussian splatting when interactive rendering and rapid scene optimization are priorities.
• Use learned reconstruction when observations are incomplete, but report uncertainty and validate inferred geometry.
• Use generative models for exploration and controlled variation, followed by technical conversion and review.
• Use hybrid systems when both visual realism and editable, physically meaningful structure are required.

Evaluation should combine several metrics. Image similarity measures such as pixel error or perceptual distance are useful but incomplete. Geometry may be assessed with depth error, point-to-surface distance, completeness, and multi-view consistency. Temporal results require flicker, identity, and motion-consistency analysis. Production assets require topology checks, memory and frame-time budgets, editability, and failure-case testing.

Illustrative Case Study: A Digital Twin for an Industrial Pump

Background

An engineering visualization team is asked to create a navigable digital twin of an industrial pump from a short handheld video and a small set of photographs. The twin will support operator training and maintenance planning. The client wants realistic appearance, measurement of selected external components, and the ability to hide covers and annotate parts.

Situation and Problem

A purely generative model can produce a visually convincing pump, but it may invent bolts, alter pipe connections, or create inconsistent views. A scene-specific radiance field can reproduce the captured viewpoints, but it does not naturally provide editable parts or reliable dimensions. The team must decide how neural methods should be combined with conventional modeling.

Analysis

The requirements distinguish appearance from structure. Novel-view realism is useful for contextual inspection, while measurement, annotation, and part isolation require explicit geometry and semantic labels. The video also contains reflective metal, moving shadows, and partially hidden surfaces, creating ambiguity. Any unobserved component must be marked as inferred rather than presented as measured fact.

Approach and Strategy

The team first calibrates cameras and extracts frames with adequate viewpoint overlap. A Gaussian-splat representation is trained to provide a fast visual reference and to identify regions requiring additional capture. Segmentation and feature matching divide the pump into major components. Visible components are reconstructed into simplified meshes using multi-view geometry and learned depth estimates, then aligned to known engineering dimensions where available.

A conventional physically based material setup is created for the mesh assets. The splat representation remains available for surrounding context, while critical components use explicit geometry. A differentiable alignment step minimizes silhouette and feature reprojection error, with dimensional measurements acting as constraints. Missing surfaces are reconstructed using engineering drawings and human review, and their provenance is recorded.

Validation includes held-out viewpoints, dimensional comparison against reference measurements, occlusion tests, annotation accuracy, and real-time performance. The interface displays confidence labels distinguishing captured, reconstructed, and inferred regions.

Outcome

The hybrid system delivers interactive visual quality while preserving editable components for training and annotation. The team does not claim that every hidden surface is a measured duplicate. Instead, uncertainty is visible in the interface and documented in the asset record. This approach requires more work than simply generating an attractive model, but it aligns the representation with the operational purpose.

Learning Points

The case demonstrates that no single neural representation satisfies all requirements. View synthesis and engineering interpretation are different tasks. Explicit constraints, provenance records, and human validation convert a plausible reconstruction into a more trustworthy professional artifact. The correct question is not which model is most impressive, but which combination of representations supports the required decisions.

Challenges, Risks, and Common Failure Modes

Data quality is a primary limitation. Sparse viewpoints create unseen regions; inaccurate camera poses cause blur or duplicated structures; inconsistent exposure confuses appearance learning; and biased datasets reduce performance on underrepresented objects or environments. Training data may also contain copyrighted material, personal information, or sensitive locations.

A second risk is hallucination. Generative systems may add plausible but unsupported details. In reconstruction, this can be mistaken for recovered truth. In content production, it can create continuity errors, incorrect text, extra fingers, broken topology, or objects that change identity between frames.

A third risk is objective mismatch. Optimizing pixel loss can produce oversmoothed results, while optimizing perceptual similarity may preserve visual impression but sacrifice geometry. An apparently successful benchmark may not predict performance on a new domain. Data leakage, inappropriate test splits, and evaluation only on favorable viewpoints can lead to inflated conclusions.

Compute and environmental cost also matter. Large models require substantial training resources, storage, and specialized hardware. Cloud inference introduces latency, availability, data-transfer, and vendor-dependence concerns. Organizations should estimate total operational cost rather than comparing only model inference time.

Common mistakes include treating a generated asset as final, ignoring camera calibration, mixing coordinate systems, failing to test novel viewpoints, using a single quality metric, omitting provenance, and allowing prompts to substitute for specifications. Another mistake is adding neural complexity where a conventional method is more predictable and easier to maintain.

Ethical, Governance, and Professional Considerations

Responsible neural graphics requires attention to consent, copyright, privacy, representation, and accountability. Facial or body reconstruction can expose biometric information. Training data and reference images may carry licenses that do not permit commercial reuse. Synthetic content can impersonate people, fabricate evidence, or obscure the boundary between documentation and invention.

Organizations should establish asset provenance, access controls, retention policies, and disclosure rules. Generated or reconstructed content used in safety, healthcare, public communication, or legal contexts should receive enhanced review. Users should know whether a visual is measured, simulated, reconstructed, or generatively invented.

Governance also concerns model behavior and auditability. Important outputs should be traceable to input data, model version, prompt or conditioning, post-processing, and reviewer decisions. Bias testing should examine relevant object categories, environments, demographic groups, and lighting conditions. Human oversight must have authority to reject or revise output; it should not be a ceremonial approval step.

Current Developments and Emerging Directions

Research is moving toward general-purpose 3D foundation models, multimodal scene understanding, dynamic and articulated radiance fields, semantic Gaussian representations, neural materials, and real-time reconstruction from mobile devices. Increasing attention is being given to 3D-consistent generation, because independent image generation does not guarantee a stable object or scene.

Another direction is the integration of neural representations with simulation. Learned models can accelerate appearance, fluid approximation, or inverse problems, while classical solvers provide physical constraints. Hybrid digital twins may combine sensor data, explicit CAD structure, and learned residual models.

Research is also addressing uncertainty estimation, calibrated confidence, test-time adaptation, compression, distillation, and energy-efficient inference. These developments are important because deployment requires more than visual quality: systems must communicate when they are uncertain and operate within predictable resource limits.

Best Practices for Professional Use

A dependable neural graphics project should begin with a clear specification and an explicit distinction between observed, estimated, and invented content. Capture should be designed for the intended task, with adequate viewpoint coverage, calibration, lighting variation, and metadata. Baselines using conventional graphics should be established before introducing a learned method.

The pipeline should preserve intermediate representations and support deterministic re-rendering where possible. Evaluation should include held-out data, difficult cases, temporal tests, resource measurements, and human review by domain experts. Outputs should be converted into the formats and budgets required by the target application rather than judged only in a research viewer.

Finally, teams should plan for failure. Confidence labels, fallback assets, manual correction tools, model monitoring, access controls, and rollback procedures make AI assistance operationally safer. Neural graphics is most effective when it augments a disciplined graphics pipeline instead of replacing specification, expertise, and verification.

Chapter Summary

AI and neural graphics extend computer graphics by learning scene representations, inverse mappings, appearance functions, and generative distributions from data. Differentiable rendering connects image evidence to optimizable scene parameters, but its gradients do not remove ambiguity or guarantee physically correct reconstruction. Neural radiance fields offer continuous view synthesis, while Gaussian splatting provides an efficient explicit set of renderable primitives. Learned reconstruction and diffusion-based generation are powerful for incomplete observations and creative variation, yet they may hallucinate unsupported content and often require multi-view, semantic, or geometric constraints.

The appropriate method depends on output requirements, observations, controllability, compute, and acceptable error. Professional pipelines should combine neural methods with conventional geometry, materials, validation, provenance, and human review. Image quality alone is insufficient; reliability, editability, temporal consistency, dimensional accuracy, licensing, privacy, and uncertainty must also be assessed. The most dependable direction is hybrid: learned components provide speed and prior knowledge, while explicit representations and domain expertise provide control and accountability.

Key Terms

Neural graphics: Computer graphics systems in which learned models represent, infer, optimize, or generate visual content.

Differentiable rendering: Rendering designed so gradients of image outputs with respect to scene or model parameters can be computed.

Inverse rendering: Estimation of geometry, materials, lighting, or camera parameters from images.

Neural radiance field: A learned continuous function that predicts density and view-dependent radiance for three-dimensional locations.

Volume rendering: Accumulation of color and opacity along a ray to form a pixel.

Gaussian splatting: Rendering a scene as projected, depth-ordered three-dimensional Gaussian primitives.

Implicit representation: A representation defined by querying a function rather than directly storing explicit surfaces.

Learned reconstruction: Data-driven inference of three-dimensional structure from incomplete observations.

View synthesis: Generation of images from viewpoints not directly observed.

Generative model: A model that learns a distribution and produces new samples resembling that distribution.

Diffusion model: A generative model trained to reverse a gradual noise-addition process.

Conditioning: Additional information, such as text, depth, or pose, used to guide generation.

Multi-view consistency: Agreement of geometry and appearance across different viewpoints.

Photometric loss: A loss based on differences between rendered and observed image values.

Semantic constraint: A requirement concerning object identity, category, parts, or meaning.

Radiance: Light energy leaving a point in a particular direction.

Hallucination: Plausible-looking content generated without adequate support in the input evidence.

Provenance: A record of the origin, processing, model, and decisions associated with an asset.

Digital twin: A digital representation of a physical object or system maintained for a defined operational purpose.

Hybrid pipeline: A workflow combining learned methods with conventional geometry, rendering, simulation, or human review.

Review Questions

Short Answer Questions

1. What distinguishes neural rendering from generative graphics?

2. Why does differentiability not guarantee correct reconstructed geometry?

3. What information does a neural radiance field predict at a sampled point?

4. State two advantages and two limitations of Gaussian splatting.

5. Why should observed, estimated, and invented content be distinguished in an AI-assisted pipeline?

Long Answer Questions

1. Explain the formulation of differentiable rendering as an optimization problem and discuss the role of loss functions and regularization.

2. Compare neural radiance fields and Gaussian splatting with respect to representation, rendering, training, editing, and failure modes.

3. Describe a complete AI-assisted content pipeline from specification and capture to validation and deployment.

4. Discuss why image quality metrics are insufficient for evaluating neural graphics systems used in professional applications.

5. Explain the ethical, governance, and provenance requirements for deploying generative visual content in an organization.

Analytical / Application Questions

1. A company needs a real-time navigable reconstruction of a static room from calibrated video, but also requires editable furniture geometry. Design a hybrid representation and justify your choice.

2. A generated 3D asset looks correct from the input camera but breaks under rotation and contains impossible topology. Diagnose the likely causes and propose a validation and remediation strategy.

3. Select a neural graphics approach for a medical visualization system in which visual plausibility is less important than dimensional reliability and traceability. Defend your decision and specify suitable safeguards.

22. Multiple Choice Questions with Answers

Question 1

Which statement best describes the central purpose of differentiable rendering?

A. To eliminate the need for camera calibration

B. To compute gradients that connect image error to scene or model parameters

C. To guarantee that reconstructed geometry is physically correct

D. To replace all conventional rendering algorithms

Correct Answer: B. To compute gradients that connect image error to scene or model parameters

Question 2

In a neural radiance field, a sampled point along a camera ray is typically associated with which quantities?

A. Only polygon indices and texture coordinates

B. Density or opacity and view-dependent color

C. A fixed lighting direction and collision response

D. Only a binary surface label

Correct Answer: B. Density or opacity and view-dependent color

Question 3

Why can a scene-specific radiance field produce visually good views while still being unsuitable for engineering measurement?

A. It cannot represent color information

B. It always requires a polygon mesh during training

C. It may reproduce appearance without recovering reliable explicit geometry

D. It can render only orthographic cameras

Correct Answer: C. It may reproduce appearance without recovering reliable explicit geometry

Question 4

What is a characteristic feature of Gaussian splatting?

A. The scene is represented by projected three-dimensional Gaussian primitives

B. Every surface must be represented by a subdivision cage

C. It generates images without using any spatial representation

D. It requires a separate manually painted image for every viewpoint

Correct Answer: A. The scene is represented by projected three-dimensional Gaussian primitives

Question 5

A single photograph is used to reconstruct the unseen back of an object. Which statement is most appropriate?

A. The hidden geometry is measured directly by the photograph

B. The reconstruction is necessarily exact if the image loss is low

C. The result is an inference based on learned priors and should carry uncertainty

D. No reconstruction is possible under any circumstances

Correct Answer: C. The result is an inference based on learned priors and should carry uncertainty

Question 6

Which requirement most strongly favors an explicit mesh over a purely implicit radiance representation?

A. Fast visual ideation from text

B. Editable topology, collision detection, and part-level manipulation

C. Memorization of a single camera view

D. Generation of arbitrary artistic styles

Correct Answer: B. Editable topology, collision detection, and part-level manipulation

Question 7

In the industrial pump case, why was a hybrid pipeline selected?

A. Neural methods cannot render reflective objects

B. The project required both visual context and editable, measurable components

C. Explicit geometry always renders faster than splats

D. Generative models cannot be conditioned on images

Correct Answer: B. The project required both visual context and editable, measurable components

Question 8

Which is a common failure caused by evaluating a generated asset only from the training viewpoint?

A. Improved multi-view consistency

B. Detection of hidden geometry errors

C. Acceptance of artifacts that become visible from novel viewpoints

D. Guaranteed dimensional accuracy

Correct Answer: C. Acceptance of artifacts that become visible from novel viewpoints

Question 9

Which practice most directly improves reproducibility and accountability in an AI-assisted graphics pipeline?

A. Deleting prompts and intermediate files after delivery

B. Recording model versions, inputs, seeds, edits, and evaluation results

C. Using only one subjective reviewer

D. Avoiding conventional baseline methods

Correct Answer: B. Recording model versions, inputs, seeds, edits, and evaluation results

Question 10

For a safety-relevant visualization system, which evaluation strategy is most appropriate?

A. Use only pixel similarity on favorable examples

B. Judge quality solely by whether users find images attractive

C. Combine geometric, multi-view, resource, uncertainty, and domain-expert validation

D. Allow the model to approve its own outputs without human review

Correct Answer: C. Combine geometric, multi-view, resource, uncertainty, and domain-expert validation