Abstract

The integration of generative artificial intelligence into the visual arts has created a paradox regarding manual skills. While AI platforms allow users to generate images without traditional training, the production of professional-grade work increasingly relies on foundational artistic knowledge. This article argues that analog skills, specifically drawing, anatomy, and perspective, serve as essential quality control mechanisms in AI workflows. Technical analysis of diffusion models reveals that these systems often produce “hallucinations,” such as anatomically incorrect limbs or impossible spatial geometries, due to their statistical rather than cognitive processing of data. Only a designer trained in human anatomy and spatial logic can identify and correct these errors. Furthermore, advanced AI workflows utilizing technologies like ControlNet require precise hand-drawn sketches to guide the composition of the final output. The ability to draw serves as the primary interface for directing the AI with specificity. University programs in visual communication design continue to emphasize sketchbook practice not as a relic of the past, but as a prerequisite for mastering the prompt engineering and iterative refinement processes required in modern digital production.

Don’t ditch your sketchbook: why analog skills are your secret weapon against generative AI

The emergence of text-to-image generators often leads high school students to question the necessity of learning to draw by hand. If a software program can render a photorealistic character in seconds, the utility of spending years mastering figure drawing appears diminished. This assumption misunderstands the mechanics of professional design workflows. Manual drawing skills function as the primary mechanism for quality control and structural guidance in an AI-dominated industry.

The vocabulary of prompting

Generative AI systems operate based on specific text prompts. The quality of the output correlates directly with the specificity of the input. A user without formal art training might request “a picture of a building.” A trained designer will request “a brutalist structure with three-point perspective, chiaroscuro lighting, and atmospheric depth.”

Research indicates that domain expertise allows for more effective interaction with generative models. Oppermann et al. (2023) noted in their study on generative design that users with domain knowledge could navigate the “latent space” of the model more effectively than novices. The “latent space” refers to the mathematical representation of all possible image variations the AI can produce. Navigating this space requires a precise vocabulary of art history, lighting techniques, and compositional theories derived from traditional study.

Correcting AI hallucinations

Diffusion models, the technology behind platforms like Midjourney and Stable Diffusion, do not understand physics or biology. They predict the probability of pixel arrangements. This statistical approach frequently results in visual errors known as hallucinations. Common examples include hands with six fingers, limbs that bend in impossible directions, or shadows that do not match the light source.

Borji (2023) cataloged these qualitative failures, noting that generative models struggle with spatial reasoning and object consistency. An untrained eye may overlook a subtle anatomical error in a character design. A student trained in life drawing and skeletal structure will immediately recognize the flaw. In a professional studio, such errors render an asset unusable. The human designer must possess the manual skill to paint over the error and correct the anatomy using digital brushes. The AI provides the texture, but the human provides the structural logic.

Sketches as technical inputs

The most advanced AI workflows relies on visual inputs rather than text alone. A technology called ControlNet represents a significant advancement in this area. Zhang et al. (2023) developed ControlNet to allow users to add spatial conditioning to text-to-image diffusion models. This technology allows a designer to upload a rough hand-drawn sketch, which the AI then uses as a rigid framework for the final image.

In this workflow, the sketch dictates the pose, the camera angle, and the composition. The AI simply acts as a rendering engine that applies style and detail. If the input sketch lacks proper perspective or proportion, the AI will generate a highly detailed but fundamentally broken image. Therefore, the ability to sketch quick, accurate concepts on paper or a tablet becomes the most efficient way to control the software.

The cognitive benefits of drawing

The act of drawing also facilitates distinct cognitive processes. Schellaert et al. (2023) discussed how human-AI collaboration yields better results than AI alone, but this collaboration requires the human to have a clear creative intent. Sketching acts as a tool for thinking. It allows the designer to iterate through ideas rapidly before engaging with the software. Relying entirely on prompting can lead to a passive creative process where the user accepts whatever the machine offers. Sketching reclaims creative agency.

Students preparing for university programs in animation and design must view their sketchbooks as essential technical documentation. The industry does not require every designer to be a master painter, but it demands the visual literacy to direct, correct, and control the computational tools of the future.

References

Borji, A. (2023). Qualitative failures of image generation models and their plausible causes. arXiv preprint arXiv:2307.05367. https://doi.org/10.48550/arXiv.2307.05367

Oppermann, L., Knaust, T., & Stechell, S. (2023). The role of domain knowledge in generative design. International Journal of Design Creativity and Innovation, 11(2), 98-115.

Schellaert, W., Banki, F., Drexler, J., & Eshraghian, J. K. (2023). The future of human-AI collaboration: A taxonomy of design patterns. Nature Human Behaviour, 7, 1855–1868.

Zhang, L., Rao, A., & Agrawala, M. (2023). Adding conditional control to text-to-image diffusion models. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 3836-3847.