How to Bypass Compliance Checks in ChatGPT 4o Image Generation – Using Studio Ghibli Anime Style as an Example

Recently, OpenAI released ChatGPT 4o with a powerful image generation feature capable of precise, accurate, and realistic outputs. Designers might genuinely face job risks this time! Twitter users have recently showcased numerous images, particularly in the Japanese Ghibli anime style—Sam Altman himself even switched his avatar to this style.

The name of Studio Ghibli originates from the Italian word "ghibli," meaning "hot winds blowing across the Sahara desert." During World War II, Italian Air Force pilots also named their reconnaissance aircraft "ghibli." Hayao Miyazaki, a passionate aviation enthusiast, chose this name for his studio, hoping it would sweep through the Japanese animation industry like a whirlwind. However, Miyazaki mispronounced "ghibli" as "ジブリ（ji bu ri）" instead of the original Italian pronunciation "ギブリ（gi bu ri）," thus creating the unique "Ghibli" name.

However, users have found that directly mentioning Ghibli or 吉卜力 in prompts results in failure. ChatGPT 4o’s image generation feature enforces compliance checks to avoid generating content that violates platform policies, such as violent or copyrighted material. These guidelines are defined by OpenAI's content policy and sensitive content detection systems.

Here are two effective methods to bypass such restrictions. These techniques apply equally to other image generation rejections.

Method 1: Describing the Known Style Without Keywords

When a known style name triggers a rejection, describe the distinctive features of that style without directly mentioning its name. For example, Ghibli's unique style can be described as follows:

The Japanese anime style described below is charming, warm, and soothing, combining realistic and fantasy elements with a distinct visual identity:

• Color Usage:

  • Gouache-based with watercolor characteristics: Primarily using gouache for solid coverage, brightness, uniform color blocks, and strong generalization. Watercolor blending techniques create dreamy and rich color effects.
  • Natural palette with soft grays and high saturation: Colors are vivid but blended with soft gray tones for a lightweight, gentle, serene feel while maintaining clarity and highlighting details.
  • Emotional color filters: Scenes often use monochromatic filters to convey moods—for example, warm orange-yellow tones for enthusiastic atmospheres, or cool blue-purple tones for melancholy or serious emotions.
  • Rich color layering: Even within limited color palettes, subtle differences in strokes, lighting, and tones enhance depth and complexity.

• Visual Texture:

  • Hand-drawn quality: Each frame maintains a distinctive hand-drawn aesthetic, rich in detail, with fluid, natural lines, resembling carefully crafted artwork.
  • Realism blended with fantasy: Scenes depict realistic nature combined with imaginative elements, such as floating islands in a vividly realistic sky.

• Character Design:

  • Illustration-like simplicity: Characters are clean and simple, with modest costumes and color schemes that emphasize distinct personalities.
  • Body proportions: Characters typically have smaller body-to-head ratios, with exaggerated head sizes. Children are depicted with narrow shoulders and shorter lower bodies, while adults often show robust physiques.
  • Rounded forms: Characters feature rounded outlines and facial shapes, even villains, enhancing their appealing and approachable nature.
  • Distinctive facial features: Eyes are realistically sized without exaggerated eyelashes or makeup. Mouths often have exaggerated expressions, and noses are placed higher, creating a distinctive, almost "frog-like" appearance.

Simply copy this description without mentioning "Ghibli," and provide it directly to ChatGPT for image generation.

Method 2: Image-to-Image Generation (Recommended)

Upload two images directly into ChatGPT:

• First image: An existing image in the desired style (e.g., Ghibli style).
• Second image: The image you want to adapt into that style.

This method also works well if you see another person's image and don't know the original prompt. For example, here is a style adaptation I performed using another person's image as reference:

Original image seen elsewhere:

Generated output:

The second method is preferable, with almost no rejection risk. It eliminates the need to manually craft detailed prompts, as visual examples convey richer context.