Image Battle

Summary for Ghibli style

This category proved to be a difficult stress test for style consistency. The "Ghibli Aesthetic" is specific: it requires a blend of simple, cel-shaded characters against complex, impressionistic watercolor backgrounds. Many top-tier models struggled to separate high fidelity from high realism.

Key Findings

• Top Performers: The Google family of models, specifically Nano Banana (2.5 Flash) and Nano Banana Pro, along with Imagen 4.0 Ultra, dominated this category. They consistently achieved near-perfect scores (9s and 10s) by accurately mimicking the 2D animation look.
• High Risk / High Reward: GPT Image 1.5 achieved a perfect 10/10 on the Howl's Moving Castle prompt but suffered from multiple safety refusals on other prompts involving specific character names.
• The 3D Trap: A significant number of models, including DALL-E 3 and Flux 1.1 Pro Ultra, frequently failed to suppress their tendency toward 3D/CGI rendering, resulting in lower scores despite high technical competency.
• Unexpected Winner: Grok Imagine performed surprisingly well, notably handling text generation better than others in the Countryside train station prompt.

Deep Dive: Mastering the Miyazaki Aesthetic

Analyzing the data across all prompts reveals distinct tiers of model performance based on their ability to replicate specific artistic nuances.

1. The "Anime vs. Realism" Struggle

The most common failure mode observed was the inability to render a true "2D" look.

• The Issue: Models like Midjourney v7 and Flux 1.1 Pro Ultra often produced images that looked like video game assets or detailed digital paintings rather than animation cels. For example, in the Ponyo Sea Creature prompt, Midjourney v7 created a "comic-book illustration" rather than the requested fluid animation style, scoring a 5/10 despite high artistic merit.
• The Solution: Models that successfully decoupled "detail" from "realism" scored highest. Nano Banana Pro consistently rendered the specific "watercolor concept art" look requested in prompts like Vegetable Garden.

2. Copyright and Safety Refusals

This category triggered significant safety protocols.

• OpenAI Models: ChatGPT 4o and GPT Image 1.5 frequently refused prompts containing names like "Kiki," "Ponyo," or "Spirited Away." However, when they did generate (e.g., Howl's Moving Castle), the results were often spectacular (10/10).
• Midjourney: Also experienced generation failures likely due to banned prompts on specific character names.

3. Text and Detail Hallucinations

• Text: In the Countryside train station prompt, many models produced gibberish on signs. Grok Imagine stood out by generating legible, correct Japanese Kanji (田舂駅 - Countryside Station), earning it a high technical score.
• Artifacts: Models like Recraft V3 occasionally hallucinated "magical creatures" where none were requested or failed to animate objects properly (e.g., static kitchens instead of magical cooking in Magical Kitchen).

4. Background vs. Foreground Separation

A key element of the Ghibli style is the visual separation between the character (cel-shaded, distinct lines) and the background (painterly, soft).

• Success: Seedream 4.0 excelled here, particularly in Totoro Meadow, scoring a 9/10 by nailing this depth-of-field effect.
• Failure: Ideogram V2 often flattened the image, making the character and background look like they were made of the same material, losing the animation feel.

Model Recommendations by Scenario

Based on the data, here are the best models for specific Ghibli-style use cases:

🎨 Best for Authentic 2D Animation Style

If you need an image that looks exactly like a screenshot from a movie:

1. Nano Banana (2.5 Flash): Consistently achieved 9s and 10s. It understands the specific color palette and line weight of 1980s-90s anime.
2. Imagen 4.0 Ultra: A powerhouse for prompt adherence. It nailed the Kiki's Delivery Service character design perfectly (10/10).
3. Nano Banana Pro: Exceptional at capturing the "concept art" feel.

🏗️ Best for Complex/Fantasy Structures

For prompts involving architecture, mechs, or steampunk elements (like Howl's Moving Castle):

1. GPT Image 1.5: Despite safety issues, when it works, it is unbeatable for intricate detail, scoring a perfect 10/10 on the castle prompt.
2. Flux 2 Pro: Performed admirably on the Nausicaä Mech prompt (9/10), capturing the rusty, insectoid aesthetic well.

🌿 Best for Nature and Backgrounds

For landscapes, gardens, and meadows:

1. Seedream 4.0: Delivers lush, atmospheric lighting. It scored a 10/10 on the Vegetable Garden prompt.
2. Ideogram 3.0 (Quality): While it struggled with creatures, it handled the Ponyo Sea Creature prompt perfectly (10/10), capturing the fluid water style better than almost any other model.

🚫 Models to Avoid for this Category

• DALL-E 3: Tendency to default to "woodblock print" or generic digital art styles when prompted for Ghibli.
• Ideogram V2: Generally scored lower (3-6 range) due to poor anatomical coherence and generic cartoon styling.
• MiniMax Image-01: Frequently ignored the 2D constraint, delivering 3D renders instead.