AI Image Style Transfer: A Creative Mashup Experiment with Ghibli and Avatar

When AI Meets Creativity: A New Approach to Style Mashups

Recently, a Twitter post sparked widespread discussion in the AI creative community. A creator showcased an experiment using AI tools to combine Singapore's urban landscape with the iconic hand-drawn animation style of Studio Ghibli, while incorporating the flying creature "Ikran" from Avatar.

What appears to be a casual experiment actually reflects the powerful capabilities of current AI image generation technology in style transfer and creative fusion. With a lighthearted tone, the creator demonstrated a bold mashup spanning three entirely different visual systems—real urban photography, Japanese animation aesthetics, and fantastical creature designs from a Hollywood sci-fi blockbuster.

The Collision of Three Visual Languages

This creative case is itself a fascinating AI style transfer experiment. The concept of Style Transfer dates back to the groundbreaking 2015 paper by Gatys et al., which used convolutional neural network (CNN) feature representations to separate image content from style. Early methods required explicit style reference images and were computationally expensive, followed by generations of improvements including fast style transfer and arbitrary style transfer. In today's era of diffusion models, style transfer no longer requires a separate technical pipeline—it's been integrated into the text-to-image generation process. After training on massive image-text pair datasets, models naturally learn to map text descriptions to corresponding visual feature spaces.

Here are the three core elements fused in this case:

Singapore's Cityscape: As a representative of modern urban landscapes, Singapore boasts a highly recognizable skyline. The sky garden connecting the three towers of Marina Bay Sands, the 50-meter-tall Supertrees at Gardens by the Bay, and the Merlion statue create a globally recognizable urban silhouette. Notably, these buildings themselves blend futuristic and natural ecological design philosophies, giving them a natural visual affinity with sci-fi or animation styles—making them particularly effective base material for AI style mashup creations.
Ghibli Style: The hand-drawn animation aesthetic created by Hayao Miyazaki and his studio is renowned for warm color palettes, delicate depictions of nature, and dreamlike atmospheres. Since Studio Ghibli's founding in 1985, under the leadership of Miyazaki and Isao Takahata, it has developed a distinctive visual aesthetic system—extensive use of soft watercolor-like gradient colors, meticulous rendering of clouds and skies, vibrant natural landscapes, and life-filled architectural details. Ghibli is also famous for refusing to rely entirely on CG technology, maintaining hand-drawn traditions. This distinctive and consistent visual language is precisely the type of style that AI models can most easily learn and reproduce.
Avatar's Ikran: The flying creatures from Pandora in James Cameron's films, brimming with alien ecological imagination. The Ikran's design blends biological forms from pterosaurs, manta rays, and tropical fish, featuring vivid fluorescent colors and translucent membrane wings—representing the pinnacle of concept design in the film industry.

Fusing these three elements means AI must understand and reconcile completely different artistic vocabularies—a level of creative difficulty that was almost unimaginable just a few years ago.

The Technical Logic Behind AI Style Transfer

This type of creation can be rapidly achieved thanks to the rapid development of diffusion model technology in recent years. Image generation tools represented by Midjourney, Stable Diffusion, and DALL-E can now combine and reconstruct different visual styles through simple text prompts.

The core idea of diffusion models originates from the diffusion process in thermodynamics. The working principle consists of two phases: the Forward Process gradually adds Gaussian noise to an image until it becomes pure random noise; the Reverse Process trains a neural network to learn how to gradually recover a clear image from noise. During actual generation, the model starts from random noise and, guided by the semantics of text prompts (achieved through text-image alignment models like CLIP), progressively denoises to generate the target image. This architecture enables the model to simultaneously respond to multiple style instructions within a single generation process, achieving natural fusion of multiple visual languages.

Why "Ghibli Style" Is a Popular Choice in AI Generation

Ghibli style has consistently been one of the most popular style tags in the AI image generation field. The reasons include:

Abundant Training Data: Miyazaki's works are widely distributed, with vast amounts of high-quality related imagery available online for models to learn from. From Nausicaä of the Valley of the Wind to Spirited Away to The Boy and the Heron, dozens of works have accumulated massive collections of screenshots, fan art, and analytical imagery worldwide.
Distinctive Style Features: Soft lighting, saturated natural colors, and hand-drawn textures—these visual characteristics are easily captured and reproduced by diffusion models. From a machine learning perspective, Ghibli style is highly clustered in feature space, allowing models to generate outputs matching the style with high confidence.
Strong Emotional Resonance: Ghibli's aesthetics evoke nostalgia and warmth in viewers, making generated results more appealing for sharing. This emotional value means related AI works spread far more efficiently on social media than other styles.

Because of this, users need only add a description like "Studio Ghibli style" to their prompts to achieve highly recognizable output.

Prompt Engineering: The Core Craft of AI Creation

Achieving this kind of complex multi-style fusion creation depends heavily on Prompt Engineering. Prompt engineering has evolved into a specialized craft within the AI image generation field, with effective prompt design requiring consideration of multiple dimensions: Subject description, Style specification, Composition requirements, Lighting conditions, and Negative Prompts to exclude unwanted elements.

In multi-style fusion scenarios, weight distribution in prompts is crucial. For example, in Stable Diffusion, users can use parentheses and numerical values to adjust each element's influence, such as (Studio Ghibli style:1.3), (Singapore skyline:1.0), (Ikran flying creature:0.8). A carefully designed prompt is essentially pinpointing the target image's coordinates in a high-dimensional semantic space, requiring the creator to have deep understanding of the model's response characteristics.

Free Combination of Cross-IP Elements

What makes this case even more noteworthy is that it freely combines visual elements from different intellectual properties (IP). AI models aren't constrained by any single work's worldbuilding—they can "transplant" Avatar's Ikran into Singapore's skies while applying a Ghibli aesthetic filter.

This creative freedom is difficult to achieve with traditional manual illustration—an artist might need days or even weeks to complete such a fusion piece, while AI image generation tools might only need a few minutes and a carefully crafted prompt. From a technical perspective, this is because diffusion models learn distributed representations of visual concepts during training. Visual elements from different IPs exist in combinable forms within the model's Latent Space, allowing them to be freely assembled like building blocks.

Creative Democratization and Copyright Concerns

How AI Tools Lower Creative Barriers

This kind of casual creative experiment is a direct manifestation of the creative democratization brought by AI tools. Creative visions that previously required professional drawing skills can now be quickly visualized by ordinary users through a few prompt sentences. This has greatly unleashed people's creativity and desire for expression.

Whether it's rapid validation of design inspiration, preliminary exploration of concept art, or purely entertaining creation, AI provides a low-barrier, high-efficiency experimental platform. It's estimated that Midjourney alone had over 16 million registered users in 2023, generating millions of images daily. This scale of creative participation is unprecedented in human art history and is redefining the boundaries of what it means to be a "creator."

The Unavoidable Copyright Controversy in AI Creation

However, this type of creation comes with ongoing controversy. Using copyrighted intellectual properties like Ghibli's style and Avatar elements in AI creation has been a hotly debated topic in the industry.

Discussions about whether AI should learn and imitate specific artists' or studios' styles have never ceased. Miyazaki himself, during a 2016 demonstration where he was shown AI-generated animated characters, stated that "this is an insult to life itself," expressing strong criticism of AI-generated animation. Whether models constitute "fair use" of protected works as training data remains a legal and ethical gray area.

As of 2024, legal frameworks regarding AI training data copyright issues are still rapidly evolving globally. In the United States, several important lawsuits are ongoing, including Getty Images v. Stability AI and class-action lawsuits by multiple artists against Midjourney and Stability AI, with the central dispute being whether using copyrighted works to train AI models constitutes "Fair Use." The EU has provided a relatively clear framework through the AI Act and the Digital Single Market Copyright Directive, allowing text and data mining but granting rights holders an "Opt-out" right. Japan has adopted a more lenient stance, with Article 30-4 of its Copyright Act allowing the use of works for AI training when the purpose is not to enjoy the works themselves. These different legal approaches will profoundly influence the future development direction of AI creative tools.

Conclusion: Industry Reflections Behind the Fun

A simple tweet actually encapsulates the current state of AI creative tool development. It demonstrates the remarkable progress of technology—making cross-style, cross-IP AI image style transfer readily accessible—while also reminding us of the deeper unresolved issues surrounding such technological applications.

For ordinary users, this is a fun-filled era of creation; for the industry as a whole, finding the balance between encouraging innovation and respecting original creators' rights remains a topic requiring ongoing exploration. When AI can so effortlessly mashup any visual style in the world, perhaps we need to seriously consider: where exactly do the boundaries of creativity lie?