피파 한 줄 정리: 'samurai' 한 단어가 ukiyo-e 색감·dramatic 조명·misty 분위기를 다 끌어와. Style attractor와 싸우려면 명시적으로 반대 방향을 밀어야 해.
Mental model: You know how someone who speaks three languages sometimes accidentally drops a French word into an English sentence? That's style leakage. The model has learned thousands of aesthetic styles from its training data, and certain keywords or concepts are so strongly associated with particular styles that they "bleed" into your image even when you didn't ask for them.
How Style Leakage Works
Every word in your prompt activates a neighborhood of learned associations. When you type "samurai warrior," you don't just activate the concept of a warrior in Japanese armor — you activate the entire aesthetic cluster: ukiyo-e woodblock print textures, cherry blossoms, dramatic ink-brush compositions, moody fog, specific color palettes. Even if you explicitly add "photorealistic" to the prompt, those learned associations still tug at the output.
"A samurai warrior in a modern office, photorealistic" → Often still gets dramatic lighting, muted cinematic color grading, and stylized atmosphere rather than mundane office lighting
"A person wearing traditional samurai armor sitting at a modern office desk, overhead fluorescent lighting, bland office environment, corporate photography style, Canon EOS R5"
Common Style Attractors
Certain keywords act as powerful style magnets that pull the entire image toward a specific aesthetic:
- "Fantasy" → pulls toward saturated colors, dramatic lighting, painterly textures
- "Cyberpunk" → neon purple/teal, rain-slick streets, Blade Runner atmosphere
- "Victorian" → sepia tones, soft focus, ornate detail
- "Anime" → cel-shading, large eyes, specific proportions, vibrant colors
- "Cinematic" → anamorphic lens distortion, shallow depth of field, color grading
These aren't bugs — they're the model faithfully reproducing correlations it learned. But they become problems when you want a fantasy scene in a photorealistic style, or an anime character in a realistic environment. The styles compete.
Aesthetic Drift Across Generations
A subtler form of drift occurs when you iterate. You generate an image, like it, use it as a reference for a variation, and repeat. Each cycle, the model may slightly amplify the dominant style, making things more saturated, more polished, more "AI-looking." This is similar to making a photocopy of a photocopy — small biases compound.
Style Conflict
Mixing contradictory style signals creates unstable results:
When style keywords conflict, the model often collapses to its default learned aesthetic — a sort of "averaged beauty" that feels generically AI-generated. This is why images from different models often have a recognizable "house style."
- Style leakage occurs because keywords activate entire aesthetic clusters, not isolated concepts.
- Certain words ("fantasy," "cyberpunk," "cinematic") are powerful style attractors.
- Mixing contradictory styles produces muddy, averaged results.
- Iterative generation can cause aesthetic convergence toward the model's defaults.
- Solution: commit to one clear style direction, use specific references, and re-anchor often.