Ghibli: A 'Diffusion' of Art & Artificial Intelligence

Disclaimer: This article was generated by GenHI - Generative Human Intelligence - and reflects my personal views (and maybe yours).

Let's start by valuing some famous and strongly opinionated lines from the originator of Studio Ghibli art-form.

“I strongly feel that this is an insult to life itself.”— Hayao Miyazaki, on watching grotesque AI-generated movement by an animation team during a 2016 NHK documentary

“I’d like to see computers and robots banished from the studio. Hand-drawn animation is what we do best.” — Hayao Miyazaki in 2009

Hello friends... As generative AI continues to evolve at fast speed, we face increasingly complex questions at the intersection of technology, creativity, and ethics. One particularly fascinating case study involves AI systems that replicate the distinctive visual aesthetic of Studio Ghibli's animated masterpieces – a style painstakingly developed over decades by director Hayao Miyazaki and his team.

The Studio Ghibli

As of March 2025, multiple AI systems can generate strikingly Ghibli-like visuals using text or image prompts. These include OpenAI's GPT-4o, fine-tuned Stable Diffusion models like those from getimg.ai, and other emerging tools offering “image-to-image” Ghibli transformations

How is this done and is this an acceptable outcome and usage of Gen AI?

Diffusion of Ethics: Artificial Intelligence versus Intelligent Art

"Is it fair for advanced AI systems to 'create' a near-replica of someone's niche, advanced, rich, hard-earned, unique, art form?"

... You mean, like someone's Python coding ability?

AI systems that mimic Studio Ghibli’s distinctive aesthetics raise questions around all creative agencies, their moral harm, and the fair treatment of original artists. The unlicensed style replication causes moral injury on creators, eroding their sense of authorship and devaluing their labor.

Legally, copyright protects specific expressions but not abstract “styles,” creating a grey zone where AI can freely learn and reproduce aesthetic patterns from the internet.

Technically, generating Ghibli‑like images relies on massive datasets scraped from the web, ubiquitous diffusion‑based pipelines, and fine‑tuning or style‑transfer modules - all of which may perpetuate opaque data practices and obscure attribution.

Studio Ghibli has expressed concerns over the use of its distinctive style in AI-generated content. Legal experts note that while mimicking a visual style isn't typically protected by copyright, using specific works without permission could raise issues. The founders have sharply criticized AI since 2016 for generating Ghibli art.

There are two main concerns here.

1. Copyright Concerns and Opaque Data Provenance:

a) Directly profiting from a style forged by decades of individual labor (as with Miyazaki’s work) edges toward exploitative appropriation - particularly when profits flow to tech firms, not original creators.

b) Current copyright law protects exact expressions (frames, characters) but not stylistic motifs. AI trainers often rely on “fair use” or public‑domain claims, but courts are increasingly scrutinizing whether scraping for style is truly non-infringing.

2. Transparency of Training Data & Informed Engagement:

a) Ethically, audiences deserve to know whether an image is human‑made, AI‑assisted, or wholly AI‑generated. This transparency respects viewer autonomy and preserves trust in artistic communication.

b) But, large‑scale crawls (e.g., LAION) amass untold millions of images without individual licensing, making it virtually impossible to trace which Ghibli frames ended up in a training set—undermining accountability.

How Does the Style Transfer Mechanics Work?

Modern generative image models like Stable Diffusion work through a process called denoising diffusion probabilistic modeling (DDPM). Here’s how this mechanism applies to Ghibli style transfer:

1. Diffusion Process:

The model is trained to reverse a process of adding noise to an image over several time steps. This way, through the iterations, the image is "de-noised" with specific bias to "unravel" the image that aligns with the requested prompt.

So, starting from pure noise, the model learns to predict and gradually recover the original image at each denoising step.

This iterative process enables very fine control over image structure, texture, and style - each of these at multiple levels of abstraction.

2. Ghibli-Style Fine-Tuning:

To generate images in the Ghibli aesthetic, the base Stable Diffusion model is simply fine-tuned on curated and actual Ghibli art collections.

Just like the traditional way, this Fine-tuning can be full (adjusting the entire model) or it can be lightweight via LoRA (Low-Rank Adaptation), which allows style adaptation by injecting a small number of additional parameters into the model, which are then trained.

I'm Pro-Tech Usually, Pro-Artist Today

Yes, learning from billions of lines of open-source code on GitHub and replacing human software engineers is eerie similar to the ethics around Ghibli we now have uproar against.

The same uproar applies to when AI has been able to generate music - some of which is uncomfortably similar to the original artists - I have often bumped into AI generated songs that taught "what if this other singer had sung it". When in fact that singer has never sung it and never will, the AI generated song sounds perfectly convincing.

We call tools like Copilot “technological advancement” or “creativity boosters” - so why is Ghibli-style generation treated differently.

1) Code is often seen as functional (solving a task), not expressive in the same way art is.

2) Reusing known design patterns or idioms is normal in programming—it’s called best practice, not plagiarism. Besides, there's the GPL licenses for code open-soured on GitHub.

The core question is the same: If an AI learns from your labor—your creativity, your years of refinement—and then generates something indistinguishable… is that okay?

A Few Technical Solutions for Perturbed Artists

As AI practitioners, we have a responsibility to develop technical approaches that respect creative labor while enabling innovation. Several promising directions include:

1. Licensing & Royalty / Revenue Share: Establish marketplaces where artists can opt in, granting AI access to their style in exchange for royalties on derivative uses. For example, we can build a blockchain-based micropayment infrastructures that compensate original artists when their styles are leveraged
2. Attribution Standards: Embed metadata tags in generated images signifying the AI model, style source, and training provenance—ensuring viewers and platforms know what they’re seeing.
3. Ethical Frameworks: Develop industry norms (akin to medical informed consent) whereby artists explicitly consent to style‑use terms before their work enters a training corpus.
4. Cryptographic Provenance: Embedding tamper-proof records of training data sources into model metadata
5. Opt-In Training Databases: Building specialized datasets that only include materials from creators who have explicitly consented
6. Style-Specific Licensing: Developing technical mechanisms that enforce licensing restrictions based on detected style characteristics
7. Automated Attribution: Creating tools that identify stylistic influences in generated images and provide appropriate credit

Open Questions for You

1) Do you have more ideas to support the artists?

2) Should AI‑style datasets only include art from creators who opt in, potentially with micro‑payments per generation?

3) Philosophically, can an artistic style ever belong to an individual, or does it live in shared human culture?

4) How do we balance the need for transparent AI systems with the practical burdens of metadata tagging and licensing compliance?

Your thoughts on these questions will help shape a future that honors both human creativity and responsible AI progress.

References

1. Vox
2. Reddit
3. The IP Press
4. Medium, Medium, Medium
5. Latest news & breaking headlines
6. UNSW Business School: BusinessThink
7. trails.umd.edu
8. Electronic Frontier Foundation
9. Michigan Tech Blogs