Advances in Artificial Intelligence: A Review for the Creative Industries

Artificial intelligence (AI) has undergone transformative advances since 2022, particularly through generative AI, large language models (LLMs), and diffusion models, fundamentally reshaping the creative industries. However, existing reviews have not comprehensively addressed these recent breakthroughs and their integrated impact across the creative production pipeline. This paper addresses this gap by providing a systematic review of AI technologies that have emerged or matured since our 2022 review, examining their applications across content creation, information analysis, post-production enhancement, compression, and quality assessment. We document how transformers, LLMs, diffusion models, and implicit neural representations have established new capabilities in text-to-image/video generation, real-time 3D reconstruction, and unified multi-task frameworks-shifting AI from support tool to core creative technology. Beyond technological advances, we analyze the trend toward unified AI frameworks that integrate multiple creative tasks, replacing task-specific solutions. We critically examine the evolving role of human-AI collaboration, where human oversight remains essential for creative direction and mitigating AI hallucinations. Finally, we identify emerging challenges including copyright concerns, bias mitigation, computational demands, and the need for robust regulatory frameworks. This review provides researchers and practitioners with a comprehensive understanding of current AI capabilities, limitations, and future trajectories in creative applications.

💡 Research Summary

The paper provides a comprehensive review of artificial intelligence advances that have reshaped the creative industries since the authors’ previous 2022 survey. It begins by outlining the rapid emergence of generative AI, large language models (LLMs), diffusion models, and implicit neural representations (INRs) between 2022 and mid‑2025. Key commercial milestones are highlighted, including OpenAI’s GPT‑3.5, GPT‑4, DALL·E 2/3, and the video‑generation model Sora; Google’s Gemini 1.5 with its expanded context window; Anthropic’s Claude 3 Opus; and the open‑source diffusion model Stable Diffusion. The authors argue that these technologies have moved AI from a peripheral support role to a core engine of content creation, analysis, post‑production, compression, and quality assessment.

Four technological pillars are examined in depth. First, transformer architectures—originally introduced for natural‑language processing—have been adapted to vision (ViT, Swin Transformer) and now dominate image and video tasks due to their global self‑attention and parallel processing capabilities. The paper also discusses the emergence of linear‑complexity state‑space models (“Mamba”) that retain transformer‑like performance with lower computational overhead. Second, LLMs built on transformers are now multimodal, handling text, images, and video. Techniques such as reinforcement learning from human feedback (RLHF) improve safety and reduce hallucinations, while prompting strategies enable nuanced creative direction. Third, diffusion models have become the de‑facto standard for high‑fidelity text‑to‑image and text‑to‑video generation. Their stochastic denoising pipelines support latent‑space manipulation, enabling real‑time 3D reconstruction and integration with neural radiance fields (NeRF) for immersive media. Fourth, INRs represent signals as continuous coordinate‑based functions, allowing lossless 3D mesh and volumetric rendering without traditional geometry pipelines.

The review maps these advances onto the creative production pipeline. Text‑to‑image/video models accelerate advertising, film pre‑visualization, and game asset creation, shrinking production cycles from weeks to hours. Unified frameworks such as “Take Painter” demonstrate how a single model can perform segmentation, low‑light enhancement, rain removal, and other post‑production tasks by treating an image pair as a prompt. AI‑enhanced video codecs are shown to approach or surpass conventional MPEG/AOM standards, though hardware constraints and lack of standardization hinder widespread adoption. Quality‑assessment models now incorporate LLMs and attention‑based architectures to predict perceptual scores with improved generalization, leveraging weakly‑supervised and unsupervised training to mitigate scarce labeled data.

Human‑AI collaboration is emphasized throughout. Prompt engineering and iterative feedback loops remain essential for steering creative intent and correcting model hallucinations. The authors stress that while AI can generate novel content, human oversight ensures artistic coherence, ethical compliance, and cultural relevance.

Finally, the paper outlines emerging challenges. Copyright and intellectual‑property questions arise as AI‑generated works blur the line between creator and tool. Bias embedded in training data can propagate discriminatory artifacts into generated media. The computational demands of trillion‑parameter models raise concerns about energy consumption, cost, and equitable access. Regulatory and standardization frameworks are still nascent, creating uncertainty for commercial deployment. The authors recommend future research directions: lightweight multimodal models for broader accessibility, robust human‑in‑the‑loop interfaces, comprehensive legal‑ethical guidelines, and standardized AI‑driven compression and quality‑assessment protocols.

In sum, this review synthesizes the state‑of‑the‑art AI technologies that have transitioned from auxiliary aids to foundational components of the creative industries, provides a critical assessment of their capabilities and limitations, and charts a roadmap for researchers and practitioners navigating the evolving landscape.