↓Skip to main content
Alemasone

AI Anime Style Photos: What You Can Actually Do

Beyond the Ghibli trend, AI generators have new limits. Learn how to write prompts and what the current copyright laws say.

7 min read Updated on
A photograph of a landscape next to its reinterpreted version in an illustrated style
On this page
In the spring of 2025, social media was flooded with photos transformed into Studio Ghibli-style illustrations using ChatGPT. A few weeks later, OpenAI tightened its rules: today, prompts that name a specific studio or author are often rejected, while describing the aesthetic you want still works perfectly. Change how you write your prompt, and you’ll get much better results.

Behind this trend is also an ongoing debate involving anyone who uses these tools: to what extent is it right to ask a model to imitate someone else’s work? Further down, I will cover both topics: how to transform a photo and write the prompt, and what the law says.

How to transform a photo, step by step #

  1. Open ChatGPT (via the app or at chatgpt.com) or Gemini, which does the same thing with its image generator. Free accounts work fine too, though they have a daily limit on the number of images.
  2. Attach the photo using the + button or the paperclip icon.
  3. Write a prompt describing the illustration you want to achieve (you’ll find how to build one below) and send it.
  4. Look at the result and ask for corrections one by one: “more warm light,” “simpler background,” “softer colors.”

If the prompt explicitly mentions Studio Ghibli, it might be rejected, or the service might automatically suggest an “anime-inspired” version without using names. This is a sign that you should describe the style instead of naming it.

What happened and where the line is drawn #

OpenAI explained that they automatically reject image prompts in the style of a living artist, whereas initially, styles from studios and broader movements were allowed. This distinction created inconsistency regarding the Ghibli case, since the studio’s founder is alive and active; following the controversy, restrictions were expanded.

From a legal standpoint, the issue is more nuanced than it seems. Style itself is not protected by copyright: no one can claim exclusive rights to a color palette, a type of line work, or a way of drawing clouds. Instead, specific works are protected through their recognizable elements: characters, scenes, and compositions.

Therefore, the line is drawn between a general aesthetic and the reproduction of elements belonging to a specific work.

A quote taken out of context. Hayao Miyazaki’s harsh statement about artificial intelligence, which went viral during this trend, actually comes from a 2016 documentary regarding a demonstration of computer-generated animation, not today’s tools. Studio Ghibli did not comment directly on the photo transformation trend, but in the fall of 2025, they were among the Japanese publishers who, through the CODA association, asked OpenAI not to use their works to train Sora 2 without authorization.

How to write a prompt that works #

The practical lesson from this whole saga is that describing works better than citing. Naming a studio often doesn’t work anymore; describing the visual elements you like always works, and it actually gives you more control than a name ever could.

A good prompt has five ingredients:

  1. The subject and what they are doing. Not just “a girl,” but “a girl running along a country path holding a straw hat.”
  2. The setting. Location, season, time of day, background elements.
  3. The lighting. This determines the atmosphere: warm late-afternoon light, overcast sky, backlighting, long shadows.
  4. The colors. Saturated greens and deep blues, pastel tones, muted colors.
  5. Technique and mood. Watercolor, hand-drawn illustration with soft outlines, digital painting; nostalgic, serene, melancholy.

A template to fill in, which works just as well in English as in Italian:

Un'illustrazione di [soggetto] in [ambientazione], [ora del giorno]
con [descrizione della luce]. Palette di [colori].
Stile: [tecnica], atmosfera [aggettivo]. [Formato o inquadratura].

You’ll often get better results in English, because it is the language most of the texts used to train these models were written in. Words like hand-drawn 2D animation, soft watercolor backgrounds, and warm afternoon light immediately convey the atmosphere you are looking for.

Correct one thing at a time. Instead of rewriting everything from scratch, generate an image, see what’s wrong, and ask for a single modification: “more warm light,” “wider shot,” “less detail in the background.” This works much better than ten prompts written from zero.

Starting from your own photo #

Generators that accept images will reinterpret them. This is what made this trend famous, and it’s also the case that requires the most care.

A starting photograph of a well-lit dog on an uncluttered background
Clear lighting and a clean background yield much better results

On the practical side, photos with a well-defined subject, clear light, and a non-busy background work best. Faces are redrawn, so they will look similar but not identical, and small details like text, hands, and distant elements remain weak points.

The same photo of the dog redesigned as an anime-style illustration with soft colors and warm light
The model doesn’t copy the photo: it redraws it, so the resemblance is loose

On the privacy side, once you upload a photo to an external service, it leaves your computer. Before doing this with photos of other people—and especially children—stop for a moment: you need the consent of the person pictured, and you should read the data processing terms at least once. Most services allow you to opt-out of having your content used for model training, but you’ll have to look for that option in the settings.

What you can do with them #

For personal use (an image for yourself, a gift, a phone wallpaper), there are no practical issues.

For public or commercial use, things change for three reasons:

  • the service’s terms of use determine what you can do with generated images, and these vary from one service to another;
  • the protection of an automatically generated image is questionable in many countries because it lacks human creative contribution; in practice, you might not be able to claim exclusive rights to it;
  • an image that is recognizably similar to a protected work remains a problem, regardless of how it was produced.

If the result is intended for a product for sale or corporate material, the prudent path is to commission an illustrator, perhaps using the generated image as a reference to explain what you want.

Alternatives to ChatGPT #

Google’s Gemini, Microsoft’s Copilot, and Midjourney also generate these types of illustrations, with different rules regarding styles that change frequently. On Midjourney, there is a model specifically designed for Japanese animation, which can be activated by adding the parameter --niji to the end of your prompt.

There are also models that run on your own computer without any filters. However, this simply shifts the responsibility to the user: just because the model is installed locally doesn’t make something illegal become legal.

FAQ #

Can you still create Ghibli-style images? #

Most major services usually reject prompts that explicitly name the studio. Describing the aesthetic (hand-drawn illustration, soft colors, countryside landscapes, warm light) still works and gives you much more control over your results.

Which version of ChatGPT do I need? #

The free version is fine, though it has a daily image limit and can be slow during peak traffic times. Subscriptions increase these limits and provide priority in the queue.

The style itself—meaning the technique and type of line work—is not: specific works and their recognizable elements, including characters, are protected. This is why the debate over generated images remains open; discussions focus more on ethics and the data used to train models rather than individual results.

Can I sell the images I generate? #

It depends on the terms of service you use, and questions remain regarding the copyright protection of an image created without human creative input: in many countries, you might not be able to claim exclusive rights. For commercial use, read the terms and conditions and, if it’s a major project, consult a professional.

Are my photos used to train the models? #

It depends on the service and your account settings. Almost all of them allow you to opt-out of having your content used for training, but this is rarely the default setting: check your settings before uploading personal photos.

Why don’t the faces ever look exactly like me? #

Because the model doesn’t copy the photo: it reinterprets it, creating a new image based on the visual description it derives from it. You always lose some resemblance, and you lose even more as the style moves further away from photography.

Read next