South American Publishing books@southam.pub 20261107T052517Z eng

9789915996653-en-1 03 01 15 9789915996653 06 10.62486/978-9915-9966-5-3 00 ED E107 00 01 01 Superintelligence Series Volume 5: Intelligent Visual Arts and Audiovisual Engineering: Generative AI, Computational Creativity and Superintelligence 1 B01 21 0009-0004-3595-4979 Emanuel Maldonado Emanuel Maldonado 1 01 eng 10 COM100000 COMPUTERS / Artificial Intelligence / Generative AI 10 ART046000 ART / Digital 10 COM016000 COMPUTERS / Artificial Intelligence / Computer Vision & Pattern Recognition 10 TEC043000 TECHNOLOGY & ENGINEERING / Television & Video 93 UYQP 20 Generative artificial intelligence; Computational creativity; Visual arts; Audiovisual engineering; Synthetic media evaluation; AI ethics and copyright 30 00 Visual production is being reshaped by systems that can synthesize images, generate video, reconstruct three-dimensional environments and assist with work after filming. These capabilities raise questions that technical performance alone cannot settle: what counts as an open model, how reliably synthetic media can be evaluated or identified, whose experiences are represented in generated imagery, and where creative and legal responsibility resides. Bringing together empirical investigations and critical reviews, the volume examines text-to-image and text-to-video models, the conditions under which they are shared, and the methods used to judge their perceptual and semantic quality. It considers the limited transfer of frequency-based detection across generators and the uneven relationship between growth in generated images in public repositories and their actual reuse. Other chapters address evaluation practices for video generation and restoration, neural radiance fields and 3D Gaussian splatting, and machine learning in editing, color grading, sound, dubbing and lip synchronization. Questions of social bias, authorship, copyright and training data connect these technical subjects to their cultural and institutional consequences. Across these areas, a shared difficulty emerges: computational measures and automated workflows can support audiovisual practice without settling questions of meaning, fairness, provenance or artistic intention. The book approaches generative visual systems as parts of a larger creative process shaped by datasets, infrastructures, professional judgment and public norms. It offers a grounded account of current methods and their limitations while clarifying the responsibilities attached to the production, evaluation and circulation of synthetic media. 04 00 Chapter 1: Authorship, copyright and creativity in machine-generated visual art; Chapter 2: Openness and licensing of text-to-image and text-to-video models on the Hugging Face Hub; Chapter 3: Classical no-reference image quality metrics and human judgments of artificially generated images; Chapter 4: Frequency-domain detectors for machine-generated images and their generalization across generators; Chapter 5: Growth, use and deletion of machine-generated images on Wikimedia Commons; Chapter 6: A systematic review of metrics and human evaluation practices for video generation models; Chapter 7: A systematic review of benchmark performance in deep learning for video super-resolution and restoration; Chapter 8: A systematic review of audit studies on social bias in text-to-image generation; Chapter 9: Neural radiance fields and 3D Gaussian splatting in audiovisual production; Chapter 10: Machine learning in post-production editing, color grading, sound, dubbing and lip synchronization 01 00 03 02 https://southam.pub/book-catalog/2026/978-9915-9966-5-3/978-9915-9966-5-3_Portada.png 1 03 06 10.62486/978-9915-9966-3-9.Ch01 01 04 1 Authorship, copyright and creativity in machine-generated visual art 1 A01 21 0000-0003-0425-7201 Carlos Alberto Gómez Cano Carlos Alberto Gómez Cano 2 A01 21 0000-0002-3669-3123 Verenice Sánchez Castillo Verenice Sánchez Castillo 01 eng 30 00 The question of who authors a work produced by a machine, and whether copyright protects it, has moved from a theoretical curiosity to a central problem of intellectual property in less than a decade. This narrative review traces how legal and philosophical treatments of authorship, copyright and creativity in AI-generated art evolved from 2018 to 2026, a window that can be divided into four periods by technical and institutional inflection points. The GAN era and the first auction sales opened the question, the diffusion models of 2021 and 2022 made it a mass concern, the institutional decisions of 2023 established a human-authorship requirement, and the period from 2024 to 2026 consolidated that requirement alongside new regulation and training-data litigation. The review synthesizes the legal and philosophical literature across these periods and finds that the law has converged on a stable human-authorship standard while the philosophy has moved toward distributed and collaborative accounts of creativity, leaving a durable gap between the two that the field has not yet closed. The trajectory is interpreted as a shift in what the debate is about, from whether machines can be authors to how much human involvement suffices and who bears responsibility for what a model was trained on. 2 03 06 10.62486/978-9915-9966-3-9.Ch02 01 04 2 Openness and licensing of text-to-image and text-to-video models on the Hugging Face Hub 1 A01 21 0000-0003-1095-1865 Yasniel Sánchez Suárez Yasniel Sánchez Suárez 01 eng 30 00 Introduction: Generative image and video models have spread rapidly across research and industry, but the meaning of their openness is contested. Empirical evidence on how licenses and documentation have evolved is scarce. Methods: Census of the Hugging Face Hub public registry filtered by the text-to-image and text-to-video pipeline tags, downloaded automatically on 2026-10-25. License categories were assigned from each model license tag and openness and documentation indicators were computed per year of creation. Results: The census identified 115,532 models, 112,574 text-to-image and 2,958 text-to-video. Text-to-image models grew from 1,454 in 2022 to 42,729 in 2025. OpenRAIL licenses were the most common (32.1%), followed by custom or unspecified licenses (38.9%) and models with no license (16.0%); permissive licenses accounted for 11.7%. The share of models with an open license and unrestricted access fell from 85.1% in 2022 to 27.1% in 2025. Library integration reached 93.0% and model cards 95.5% of models. Conclusions: The volume of open image and video models expanded several-fold while their licensing shifted away from unrestricted open terms toward use-restricted and underspecified licenses. Documentation is near universal, but licensing openness has narrowed over time. 3 03 06 10.62486/978-9915-9966-3-9.Ch03 01 04 3 Classical no-reference image quality metrics and human judgments of artificially generated images 1 A01 21 0009-0002-2995-7028 Héctor Aaron Arellano Cruz Héctor Aaron Arellano Cruz 2 A01 21 0009-0008-2647-9386 Víctor Manuel Martínez García Víctor Manuel Martínez García 3 A01 21 0009-0009-7232-9749 Yennifer Díaz Romero Yennifer Díaz Romero 4 A01 21 0009-0002-7802-921X Celso Fernando Benitez de La Herrán Celso Fernando Benitez de La Herrán 5 A01 21 0009-0005-1631-4711 Rosa Edilma Garzón González Rosa Edilma Garzón González 6 A01 21 0000-0001-7536-8588 Alma delia Figueroa Suarez Alma delia Figueroa Suarez 01 eng 30 00 Introduction: Artificially generated images are produced in large volumes, but their perceptual quality varies widely and is expensive to measure with human raters. Classical no-reference metrics are a cheap substitute whose validity for generated images is uncertain. Methods: A random sample of 600 images from the AGIQA-3K database was drawn with a fixed seed. The no-reference metrics BRISQUE, NIQE and PIQE were computed for each image and compared with the published mean opinion scores for perceived quality and for text-to-image alignment using Pearson and Spearman correlations with 95% confidence intervals. Results: All three metrics correlated negatively with perceived quality, as expected, at magnitudes between -0.317 and -0.487. NIQE showed the strongest association (r = -0.49) and PIQE the weakest (r = -0.32). Correlations with text-to-image alignment were lower, ranging from -0.239 to -0.370. The mean opinion scores for quality and alignment correlated at r = 0.824. Conclusions: Classical no-reference metrics explain only a moderate share of the variance in human judgments of generated images, with NIQE performing best among the three. They should be treated as coarse quality screens rather than substitutes for human evaluation of generated content. 4 03 06 10.62486/978-9915-9966-3-9.Ch04 01 04 4 Frequency-domain detectors for machine-generated images and their generalization across generators 1 A01 21 0000-0002-3348-2267 Vugar Hacimahmud Abdullayev Vugar Hacimahmud Abdullayev 01 eng 30 00 Introduction: Machine-generated images are now easy to produce and hard to distinguish from photographs, and reliable detection is needed for forensics and content moderation. Deep detectors work well on familiar generators but degrade on unseen ones; cheap frequency-domain features are an alternative whose cross-generator behavior is unclear. Methods: A sample of 500 real photographs and 500 images from each of three generators (BigGAN, GLIDE, VQDM) was drawn from the GenImage benchmark. Fifty-nine frequency-domain features were extracted from each image and fed to a logistic regression trained on one generator, then evaluated in-domain and on the two held-out generators with accuracy and area under the curve. Results: Mean in-domain accuracy was 0.77 and fell to 0.65 across generators, an average drop of 12 percentage points. The drop was uneven, with GLIDE reaching 0.92 in-domain while most cross-generator transfers fell below 0.66, and one transfer was weaker than a coin toss. Conclusions: Frequency-domain features support lightweight detectors that work within a generator but transfer poorly to unseen generators, because they capture generator-specific spectral artifacts. They are a useful first-line screen when a detector must run without a learned model, but not a substitute for generators-diverse training. 5 03 06 10.62486/978-9915-9966-3-9.Ch05 01 04 5 Growth, use and deletion of machine-generated images on Wikimedia Commons 1 A01 21 0000-0002-5514-2992 Matilde Bolaño-García Matilde Bolaño-García 01 eng 30 00 Introduction: Wikimedia Commons, the free media repository of Wikipedia, has absorbed a rising number of machine-generated images, and their place in the project is contested. Systematic evidence on their growth, use and deletion is lacking. Methods: A registry census was conducted of the Commons category tree for machine-generated images using the MediaWiki API. All file members of the category and its subcategories were enumerated, and their upload timestamps, global usage and membership in quality-flag categories were recorded on a single snapshot date. Results: The census identified 2,548 image files across 104 subcategories. Uploads grew from 4 files in 2019 to 164 in 2022, 324 in 2024, and 1,035 in 2026. Only 2.0% of files were used in at least one wiki page. Membership in quality-flag categories reached 0.7% for low-quality tagging and 0.9% for obvious artifacts. Conclusions: Machine-generated images on Commons grew rapidly but remain almost entirely unused, and only a small share are flagged for deletion-related concerns. The repository is absorbing generated content far faster than the project is integrating it. 6 03 06 10.62486/978-9915-9966-3-9.Ch06 01 04 6 A systematic review of metrics and human evaluation practices for video generation models 1 A01 21 0000-0002-6667-7976 Rocío de la Cruz Hernández Rocío de la Cruz Hernández 01 eng 30 00 Introduction: Video generation models have advanced rapidly, but the way they are evaluated has not kept pace, and the field lacks a synthesis of which automatic metrics and human-evaluation protocols are actually used. Methods: A systematic review was conducted following PRISMA 2020. Four bibliographic sources were searched for empirical studies of video generation evaluation published between 2018 and 2026. Records were screened on title and abstract and then on full text, and automatic metrics and human-evaluation protocols were extracted from the included studies. Results: The search identified 210 records, of which 11 studies were included. All included studies reported human evaluation, and all reported the frame-level Fréchet inception distance, while the video-level Fréchet video distance was used in 72.7% of studies and CLIP-based text-alignment scores in 45.5%. Human-evaluation protocols varied in rater pool, scale and instructions, and few studies reported these details completely. Conclusions: Evaluation of video generation rests on a small set of inherited image metrics plus universal but inconsistently reported human studies. The field would benefit from standardized reporting of both metric choices and human-evaluation protocols. 7 03 06 10.62486/978-9915-9966-3-9.Ch07 01 04 7 A systematic review of benchmark performance in deep learning for video super-resolution and restoration 1 A01 21 0000-0003-2581-1873 Rolando Eslava Zapata Rolando Eslava Zapata 2 A01 21 0009-0008-6878-0441 Rómulo Esteban Montilla Rómulo Esteban Montilla 3 A01 21 0009-0002-0970-2810 Edixon Chacón Guerrero Edixon Chacón Guerrero 01 eng 30 00 Introduction: Deep learning has become the standard approach for video super-resolution and restoration, but the field lacks a systematic account of how these methods perform on standard benchmarks and how architectures compare in quality and efficiency. Methods: A systematic review was conducted following PRISMA 2020. Bibliographic sources were searched for empirical studies of deep learning video super-resolution and restoration published between 2018 and 2026. Benchmark datasets, quality metrics, and reported performance were extracted from the included studies and summarized as an evidence matrix. Results: The search identified 240 records and included 12 studies. Peak signal-to-noise ratio was reported in every study and the structural similarity index in 91.7%. The most used benchmarks were Vid4 (75.0% of studies), REDS (66.7%), Vimeo-90K (50.0%) and UDM10 (41.7%). Reported performance concentrated on a narrow set of benchmarks, and efficiency was reported far less often than quality. Conclusions: Deep learning video super-resolution and restoration is evaluated almost exclusively by two quality metrics on a small set of benchmarks, with efficiency largely unmeasured. The field would benefit from broader benchmark coverage and routine reporting of computational cost. 8 03 06 10.62486/978-9915-9966-3-9.Ch08 01 04 8 A systematic review of audit studies on social bias in text-to-image generation 1 A01 21 0009-0006-5159-9975 Gabino Ochoa Lozano Gabino Ochoa Lozano 2 A01 21 0009-0008-2647-9386 Víctor Manuel Martínez García Víctor Manuel Martínez García 3 A01 21 0009-0009-7232-9749 Yennifer Díaz Romero Yennifer Díaz Romero 4 A01 21 0009-0004-6321-4054 Yadira del Carmen Coronado Toto Yadira del Carmen Coronado Toto 5 A01 21 0000-0001-7536-8588 Alma delia Figueroa Suarez Alma delia Figueroa Suarez 6 A01 21 0009-0000-6273-071X José de Jesús Reyes Machain José de Jesús Reyes Machain 01 eng 30 00 Introduction: Text-to-image models are widely deployed, and audit studies have repeatedly shown that they reproduce gender, racial and cultural stereotypes, but the field lacks a systematic account of what those studies measure and what they find. Methods: A systematic review was conducted following PRISMA 2020. Bibliographic sources were searched for empirical audit studies of social bias in text-to-image models published between 2018 and 2026. Bias dimensions, measurement methods, documented biases and mitigation strategies were extracted from the included studies. Results: The search identified 200 records and included 11 studies. Gender and racial bias were audited in every study, and cultural or national bias in the same proportion, while mitigation strategies were evaluated in most. The documented biases were consistent in direction, with occupations and social roles persistently stereotyped along gender and racial lines. Conclusions: Audit studies of text-to-image generation converge on the finding that these models reproduce gendered, racialized and culturally skewed representations, and mitigation is discussed more often than it is rigorously evaluated. The field needs standardized audit protocols and evaluation of mitigations at scale. 9 03 06 10.62486/978-9915-9966-3-9.Ch09 01 04 9 Neural radiance fields and 3D Gaussian splatting in audiovisual production 1 A01 21 0009-0009-5129-4970 Fiorella Luz De La Cruz Espinoza Fiorella Luz De La Cruz Espinoza 2 A01 21 0000-0002-0631-5906 Norma Beatriz Ríos Norma Beatriz Ríos 3 A01 21 0000-0003-3901-6117 Carlos Jesús Canova Barrios Carlos Jesús Canova Barrios 01 eng 30 00 Neural rendering has introduced two dominant approaches to synthesizing photorealistic views of a scene, neural radiance fields and 3D Gaussian splatting, whose different trade-offs matter directly for audiovisual production. This narrative review synthesizes the two approaches across five axes, representation and training, rendering quality, rendering speed, editability, and dynamic-scene support, drawing on the literature published between 2018 and 2026. Neural radiance fields represent a scene as a continuous implicit function optimized by volume rendering, which yields high quality and strong anti-aliasing but is slow to train and render without baking or acceleration. Three-dimensional Gaussian splatting represents the same scene as an explicit set of anisotropic Gaussians rendered by rasterization, which delivers real-time frame rates with comparable quality but introduces artifacts and editing challenges of its own. The two approaches also differ in editability and in their treatment of moving scenes. The synthesis concludes that the choice between them is a production decision rather than a technical ranking, and that the field's open problems, temporal consistency, production-grade editability and standardized evaluation, are the likely routes by which neural rendering will enter routine audiovisual workflows. 10 03 06 10.62486/978-9915-9966-3-9.Ch10 01 04 10 Machine learning in post-production editing, color grading, sound, dubbing and lip synchronization 1 A01 21 0009-0004-3595-4979 Emanuel Maldonado Emanuel Maldonado 01 eng 30 00 Post-production is the stage at which a film or video becomes coherent, and machine learning has begun to change each of its core tasks in turn. This narrative review synthesizes how machine learning has advanced five post-production tasks, editing, color grading, sound design, dubbing and lip synchronization, drawing on the literature published between 2018 and 2026. Editing has moved toward assisted organization, shot selection and transcript-driven assembly; color grading toward deep colorization, exemplar-based transfer and temporally consistent stylization; sound design toward source separation, localization and generative audio; dubbing toward voice conversion, speech translation and multilingual systems; and lip synchronization toward diffusion-based and editing-based talking-head generation. The review finds that the five tasks have advanced unevenly, with lip synchronization the most active and color grading the least documented in the scholarly literature, and that a common set of limits, temporal consistency, identity preservation, creative control and evaluation, recurs across the tasks. It concludes that machine learning has matured into a reliable assistant for post-production but has not replaced the creative decisions that the discipline still reserves for human practitioners. South American Publishing 01 South American Publishing 02 https://southam.pub/book-catalog/isbn/978-9915-9966-5-3.php Montevideo UY 04 01 20261230 01 WORLD 06 15 9789915996646 9789915996646-en-2 03 01 15 9789915996646 06 10.62486/978-9915-9966-5-3 00 BC 01 01 Superintelligence Series Volume 5: Intelligent Visual Arts and Audiovisual Engineering: Generative AI, Computational Creativity and Superintelligence 1 B01 21 0009-0004-3595-4979 Emanuel Maldonado Emanuel Maldonado 1 01 eng 10 COM100000 COMPUTERS / Artificial Intelligence / Generative AI 10 ART046000 ART / Digital 10 COM016000 COMPUTERS / Artificial Intelligence / Computer Vision & Pattern Recognition 10 TEC043000 TECHNOLOGY & ENGINEERING / Television & Video 93 UYQP 20 Generative artificial intelligence; Computational creativity; Visual arts; Audiovisual engineering; Synthetic media evaluation; AI ethics and copyright 30 00 Visual production is being reshaped by systems that can synthesize images, generate video, reconstruct three-dimensional environments and assist with work after filming. These capabilities raise questions that technical performance alone cannot settle: what counts as an open model, how reliably synthetic media can be evaluated or identified, whose experiences are represented in generated imagery, and where creative and legal responsibility resides. Bringing together empirical investigations and critical reviews, the volume examines text-to-image and text-to-video models, the conditions under which they are shared, and the methods used to judge their perceptual and semantic quality. It considers the limited transfer of frequency-based detection across generators and the uneven relationship between growth in generated images in public repositories and their actual reuse. Other chapters address evaluation practices for video generation and restoration, neural radiance fields and 3D Gaussian splatting, and machine learning in editing, color grading, sound, dubbing and lip synchronization. Questions of social bias, authorship, copyright and training data connect these technical subjects to their cultural and institutional consequences. Across these areas, a shared difficulty emerges: computational measures and automated workflows can support audiovisual practice without settling questions of meaning, fairness, provenance or artistic intention. The book approaches generative visual systems as parts of a larger creative process shaped by datasets, infrastructures, professional judgment and public norms. It offers a grounded account of current methods and their limitations while clarifying the responsibilities attached to the production, evaluation and circulation of synthetic media. 04 00 Chapter 1: Authorship, copyright and creativity in machine-generated visual art; Chapter 2: Openness and licensing of text-to-image and text-to-video models on the Hugging Face Hub; Chapter 3: Classical no-reference image quality metrics and human judgments of artificially generated images; Chapter 4: Frequency-domain detectors for machine-generated images and their generalization across generators; Chapter 5: Growth, use and deletion of machine-generated images on Wikimedia Commons; Chapter 6: A systematic review of metrics and human evaluation practices for video generation models; Chapter 7: A systematic review of benchmark performance in deep learning for video super-resolution and restoration; Chapter 8: A systematic review of audit studies on social bias in text-to-image generation; Chapter 9: Neural radiance fields and 3D Gaussian splatting in audiovisual production; Chapter 10: Machine learning in post-production editing, color grading, sound, dubbing and lip synchronization 01 00 03 02 https://southam.pub/book-catalog/2026/978-9915-9966-5-3/978-9915-9966-5-3_Portada.png 1 03 06 10.62486/978-9915-9966-3-9.Ch01 01 04 1 Authorship, copyright and creativity in machine-generated visual art 1 A01 21 0000-0003-0425-7201 Carlos Alberto Gómez Cano Carlos Alberto Gómez Cano 2 A01 21 0000-0002-3669-3123 Verenice Sánchez Castillo Verenice Sánchez Castillo 01 eng 30 00 The question of who authors a work produced by a machine, and whether copyright protects it, has moved from a theoretical curiosity to a central problem of intellectual property in less than a decade. This narrative review traces how legal and philosophical treatments of authorship, copyright and creativity in AI-generated art evolved from 2018 to 2026, a window that can be divided into four periods by technical and institutional inflection points. The GAN era and the first auction sales opened the question, the diffusion models of 2021 and 2022 made it a mass concern, the institutional decisions of 2023 established a human-authorship requirement, and the period from 2024 to 2026 consolidated that requirement alongside new regulation and training-data litigation. The review synthesizes the legal and philosophical literature across these periods and finds that the law has converged on a stable human-authorship standard while the philosophy has moved toward distributed and collaborative accounts of creativity, leaving a durable gap between the two that the field has not yet closed. The trajectory is interpreted as a shift in what the debate is about, from whether machines can be authors to how much human involvement suffices and who bears responsibility for what a model was trained on. 2 03 06 10.62486/978-9915-9966-3-9.Ch02 01 04 2 Openness and licensing of text-to-image and text-to-video models on the Hugging Face Hub 1 A01 21 0000-0003-1095-1865 Yasniel Sánchez Suárez Yasniel Sánchez Suárez 01 eng 30 00 Introduction: Generative image and video models have spread rapidly across research and industry, but the meaning of their openness is contested. Empirical evidence on how licenses and documentation have evolved is scarce. Methods: Census of the Hugging Face Hub public registry filtered by the text-to-image and text-to-video pipeline tags, downloaded automatically on 2026-10-25. License categories were assigned from each model license tag and openness and documentation indicators were computed per year of creation. Results: The census identified 115,532 models, 112,574 text-to-image and 2,958 text-to-video. Text-to-image models grew from 1,454 in 2022 to 42,729 in 2025. OpenRAIL licenses were the most common (32.1%), followed by custom or unspecified licenses (38.9%) and models with no license (16.0%); permissive licenses accounted for 11.7%. The share of models with an open license and unrestricted access fell from 85.1% in 2022 to 27.1% in 2025. Library integration reached 93.0% and model cards 95.5% of models. Conclusions: The volume of open image and video models expanded several-fold while their licensing shifted away from unrestricted open terms toward use-restricted and underspecified licenses. Documentation is near universal, but licensing openness has narrowed over time. 3 03 06 10.62486/978-9915-9966-3-9.Ch03 01 04 3 Classical no-reference image quality metrics and human judgments of artificially generated images 1 A01 21 0009-0002-2995-7028 Héctor Aaron Arellano Cruz Héctor Aaron Arellano Cruz 2 A01 21 0009-0008-2647-9386 Víctor Manuel Martínez García Víctor Manuel Martínez García 3 A01 21 0009-0009-7232-9749 Yennifer Díaz Romero Yennifer Díaz Romero 4 A01 21 0009-0002-7802-921X Celso Fernando Benitez de La Herrán Celso Fernando Benitez de La Herrán 5 A01 21 0009-0005-1631-4711 Rosa Edilma Garzón González Rosa Edilma Garzón González 6 A01 21 0000-0001-7536-8588 Alma delia Figueroa Suarez Alma delia Figueroa Suarez 01 eng 30 00 Introduction: Artificially generated images are produced in large volumes, but their perceptual quality varies widely and is expensive to measure with human raters. Classical no-reference metrics are a cheap substitute whose validity for generated images is uncertain. Methods: A random sample of 600 images from the AGIQA-3K database was drawn with a fixed seed. The no-reference metrics BRISQUE, NIQE and PIQE were computed for each image and compared with the published mean opinion scores for perceived quality and for text-to-image alignment using Pearson and Spearman correlations with 95% confidence intervals. Results: All three metrics correlated negatively with perceived quality, as expected, at magnitudes between -0.317 and -0.487. NIQE showed the strongest association (r = -0.49) and PIQE the weakest (r = -0.32). Correlations with text-to-image alignment were lower, ranging from -0.239 to -0.370. The mean opinion scores for quality and alignment correlated at r = 0.824. Conclusions: Classical no-reference metrics explain only a moderate share of the variance in human judgments of generated images, with NIQE performing best among the three. They should be treated as coarse quality screens rather than substitutes for human evaluation of generated content. 4 03 06 10.62486/978-9915-9966-3-9.Ch04 01 04 4 Frequency-domain detectors for machine-generated images and their generalization across generators 1 A01 21 0000-0002-3348-2267 Vugar Hacimahmud Abdullayev Vugar Hacimahmud Abdullayev 01 eng 30 00 Introduction: Machine-generated images are now easy to produce and hard to distinguish from photographs, and reliable detection is needed for forensics and content moderation. Deep detectors work well on familiar generators but degrade on unseen ones; cheap frequency-domain features are an alternative whose cross-generator behavior is unclear. Methods: A sample of 500 real photographs and 500 images from each of three generators (BigGAN, GLIDE, VQDM) was drawn from the GenImage benchmark. Fifty-nine frequency-domain features were extracted from each image and fed to a logistic regression trained on one generator, then evaluated in-domain and on the two held-out generators with accuracy and area under the curve. Results: Mean in-domain accuracy was 0.77 and fell to 0.65 across generators, an average drop of 12 percentage points. The drop was uneven, with GLIDE reaching 0.92 in-domain while most cross-generator transfers fell below 0.66, and one transfer was weaker than a coin toss. Conclusions: Frequency-domain features support lightweight detectors that work within a generator but transfer poorly to unseen generators, because they capture generator-specific spectral artifacts. They are a useful first-line screen when a detector must run without a learned model, but not a substitute for generators-diverse training. 5 03 06 10.62486/978-9915-9966-3-9.Ch05 01 04 5 Growth, use and deletion of machine-generated images on Wikimedia Commons 1 A01 21 0000-0002-5514-2992 Matilde Bolaño-García Matilde Bolaño-García 01 eng 30 00 Introduction: Wikimedia Commons, the free media repository of Wikipedia, has absorbed a rising number of machine-generated images, and their place in the project is contested. Systematic evidence on their growth, use and deletion is lacking. Methods: A registry census was conducted of the Commons category tree for machine-generated images using the MediaWiki API. All file members of the category and its subcategories were enumerated, and their upload timestamps, global usage and membership in quality-flag categories were recorded on a single snapshot date. Results: The census identified 2,548 image files across 104 subcategories. Uploads grew from 4 files in 2019 to 164 in 2022, 324 in 2024, and 1,035 in 2026. Only 2.0% of files were used in at least one wiki page. Membership in quality-flag categories reached 0.7% for low-quality tagging and 0.9% for obvious artifacts. Conclusions: Machine-generated images on Commons grew rapidly but remain almost entirely unused, and only a small share are flagged for deletion-related concerns. The repository is absorbing generated content far faster than the project is integrating it. 6 03 06 10.62486/978-9915-9966-3-9.Ch06 01 04 6 A systematic review of metrics and human evaluation practices for video generation models 1 A01 21 0000-0002-6667-7976 Rocío de la Cruz Hernández Rocío de la Cruz Hernández 01 eng 30 00 Introduction: Video generation models have advanced rapidly, but the way they are evaluated has not kept pace, and the field lacks a synthesis of which automatic metrics and human-evaluation protocols are actually used. Methods: A systematic review was conducted following PRISMA 2020. Four bibliographic sources were searched for empirical studies of video generation evaluation published between 2018 and 2026. Records were screened on title and abstract and then on full text, and automatic metrics and human-evaluation protocols were extracted from the included studies. Results: The search identified 210 records, of which 11 studies were included. All included studies reported human evaluation, and all reported the frame-level Fréchet inception distance, while the video-level Fréchet video distance was used in 72.7% of studies and CLIP-based text-alignment scores in 45.5%. Human-evaluation protocols varied in rater pool, scale and instructions, and few studies reported these details completely. Conclusions: Evaluation of video generation rests on a small set of inherited image metrics plus universal but inconsistently reported human studies. The field would benefit from standardized reporting of both metric choices and human-evaluation protocols. 7 03 06 10.62486/978-9915-9966-3-9.Ch07 01 04 7 A systematic review of benchmark performance in deep learning for video super-resolution and restoration 1 A01 21 0000-0003-2581-1873 Rolando Eslava Zapata Rolando Eslava Zapata 2 A01 21 0009-0008-6878-0441 Rómulo Esteban Montilla Rómulo Esteban Montilla 3 A01 21 0009-0002-0970-2810 Edixon Chacón Guerrero Edixon Chacón Guerrero 01 eng 30 00 Introduction: Deep learning has become the standard approach for video super-resolution and restoration, but the field lacks a systematic account of how these methods perform on standard benchmarks and how architectures compare in quality and efficiency. Methods: A systematic review was conducted following PRISMA 2020. Bibliographic sources were searched for empirical studies of deep learning video super-resolution and restoration published between 2018 and 2026. Benchmark datasets, quality metrics, and reported performance were extracted from the included studies and summarized as an evidence matrix. Results: The search identified 240 records and included 12 studies. Peak signal-to-noise ratio was reported in every study and the structural similarity index in 91.7%. The most used benchmarks were Vid4 (75.0% of studies), REDS (66.7%), Vimeo-90K (50.0%) and UDM10 (41.7%). Reported performance concentrated on a narrow set of benchmarks, and efficiency was reported far less often than quality. Conclusions: Deep learning video super-resolution and restoration is evaluated almost exclusively by two quality metrics on a small set of benchmarks, with efficiency largely unmeasured. The field would benefit from broader benchmark coverage and routine reporting of computational cost. 8 03 06 10.62486/978-9915-9966-3-9.Ch08 01 04 8 A systematic review of audit studies on social bias in text-to-image generation 1 A01 21 0009-0006-5159-9975 Gabino Ochoa Lozano Gabino Ochoa Lozano 2 A01 21 0009-0008-2647-9386 Víctor Manuel Martínez García Víctor Manuel Martínez García 3 A01 21 0009-0009-7232-9749 Yennifer Díaz Romero Yennifer Díaz Romero 4 A01 21 0009-0004-6321-4054 Yadira del Carmen Coronado Toto Yadira del Carmen Coronado Toto 5 A01 21 0000-0001-7536-8588 Alma delia Figueroa Suarez Alma delia Figueroa Suarez 6 A01 21 0009-0000-6273-071X José de Jesús Reyes Machain José de Jesús Reyes Machain 01 eng 30 00 Introduction: Text-to-image models are widely deployed, and audit studies have repeatedly shown that they reproduce gender, racial and cultural stereotypes, but the field lacks a systematic account of what those studies measure and what they find. Methods: A systematic review was conducted following PRISMA 2020. Bibliographic sources were searched for empirical audit studies of social bias in text-to-image models published between 2018 and 2026. Bias dimensions, measurement methods, documented biases and mitigation strategies were extracted from the included studies. Results: The search identified 200 records and included 11 studies. Gender and racial bias were audited in every study, and cultural or national bias in the same proportion, while mitigation strategies were evaluated in most. The documented biases were consistent in direction, with occupations and social roles persistently stereotyped along gender and racial lines. Conclusions: Audit studies of text-to-image generation converge on the finding that these models reproduce gendered, racialized and culturally skewed representations, and mitigation is discussed more often than it is rigorously evaluated. The field needs standardized audit protocols and evaluation of mitigations at scale. 9 03 06 10.62486/978-9915-9966-3-9.Ch09 01 04 9 Neural radiance fields and 3D Gaussian splatting in audiovisual production 1 A01 21 0009-0009-5129-4970 Fiorella Luz De La Cruz Espinoza Fiorella Luz De La Cruz Espinoza 2 A01 21 0000-0002-0631-5906 Norma Beatriz Ríos Norma Beatriz Ríos 3 A01 21 0000-0003-3901-6117 Carlos Jesús Canova Barrios Carlos Jesús Canova Barrios 01 eng 30 00 Neural rendering has introduced two dominant approaches to synthesizing photorealistic views of a scene, neural radiance fields and 3D Gaussian splatting, whose different trade-offs matter directly for audiovisual production. This narrative review synthesizes the two approaches across five axes, representation and training, rendering quality, rendering speed, editability, and dynamic-scene support, drawing on the literature published between 2018 and 2026. Neural radiance fields represent a scene as a continuous implicit function optimized by volume rendering, which yields high quality and strong anti-aliasing but is slow to train and render without baking or acceleration. Three-dimensional Gaussian splatting represents the same scene as an explicit set of anisotropic Gaussians rendered by rasterization, which delivers real-time frame rates with comparable quality but introduces artifacts and editing challenges of its own. The two approaches also differ in editability and in their treatment of moving scenes. The synthesis concludes that the choice between them is a production decision rather than a technical ranking, and that the field's open problems, temporal consistency, production-grade editability and standardized evaluation, are the likely routes by which neural rendering will enter routine audiovisual workflows. 10 03 06 10.62486/978-9915-9966-3-9.Ch10 01 04 10 Machine learning in post-production editing, color grading, sound, dubbing and lip synchronization 1 A01 21 0009-0004-3595-4979 Emanuel Maldonado Emanuel Maldonado 01 eng 30 00 Post-production is the stage at which a film or video becomes coherent, and machine learning has begun to change each of its core tasks in turn. This narrative review synthesizes how machine learning has advanced five post-production tasks, editing, color grading, sound design, dubbing and lip synchronization, drawing on the literature published between 2018 and 2026. Editing has moved toward assisted organization, shot selection and transcript-driven assembly; color grading toward deep colorization, exemplar-based transfer and temporally consistent stylization; sound design toward source separation, localization and generative audio; dubbing toward voice conversion, speech translation and multilingual systems; and lip synchronization toward diffusion-based and editing-based talking-head generation. The review finds that the five tasks have advanced unevenly, with lip synchronization the most active and color grading the least documented in the scholarly literature, and that a common set of limits, temporal consistency, identity preservation, creative control and evaluation, recurs across the tasks. It concludes that machine learning has matured into a reliable assistant for post-production but has not replaced the creative decisions that the discipline still reserves for human practitioners. South American Publishing 01 South American Publishing 02 https://southam.pub/book-catalog/isbn/978-9915-9966-5-3.php Montevideo UY 04 01 20261230 01 WORLD 06 15 9789915996653