Image Battle

Summary for Photorealistic People & Portraits

This analysis reveals a distinct tier list in the capability of AI models to generate convincing human subjects. The dataset highlights a battle between stylized perfection and gritty realism.

Key Findings

• Texture is King: The highest-scoring models, such as Midjourney v7 and Nano Banana Pro, excelled because they dared to render "imperfect" skin. Older or less tuned models often failed by smoothing skin into a plastic, doll-like surface.
• The Uncanny Valley Exists: The Toddler Portrait prompt was the most significant stumbling block. Many models produced doll-like or age-inappropriate features, causing scores to plummet.
• Lighting Mastery: Models like ChatGPT 4o and Ideogram 3.0 (Quality) achieved perfect scores in complex lighting scenarios, such as the Neon Night Portrait, proving that lighting is just as critical as texture for realism.
• Top Performers: Midjourney v7, Recraft V3, and Grok 2 Image consistently delivered top-tier results across diverse ages and settings.

Deep Dive: Patterns & Performance

Across the dataset, several distinct patterns emerged regarding model strengths and weaknesses.

1. The "Plastic Skin" Problem

A recurring issue in lower-scoring models is the inability to render organic skin texture.

• Weakness: Models like DALL-E 3 and Z-Image Turbo frequently received deductions for "waxy" or "airbrushed" skin. For example, in the Elderly Portrait, DALL-E 3 scored a 5/10 due to "physically implausible reflections" and synthetic textures.
• Strength: In contrast, Midjourney v7 and Grok 2 Image scored perfect 10s in prompts like Facial Tattoos by rendering pores, stubble, and ink fade with hyper-realistic precision.

2. Anatomical coherence vs. Complexity

Models generally handled faces well but struggled when secondary anatomy interacted with props.

• The Fisherman Portrait revealed significant issues with hand anatomy for models like Ideogram V2 and Nano Banana Pro (in one instance), where fingers merged with the pipe or looked gnarled.
• However, Nano Banana Pro redeemed itself in the Freckled B&W Portrait with a perfect 10/10, showing that close-ups are often safer than mid-shots for some engines.

3. Text and Environmental Details

Background realism played a huge role in the Neon Night Portrait.

• Failure Mode: Flux 1.1 Pro Ultra and Recraft V3 struggled with gibberish text on neon signs, breaking immersion.
• Success: Ideogram 3.0 (Quality) and Grok Imagine rendered legible, coherent background text (e.g., "RAMEN", "OPEN 24H"), elevating the overall realism score significantly.

4. The Challenge of Youth

Generating the Toddler Portrait proved exceptionally difficult.

• Many models, including Flux 1.1 Pro Ultra and Midjourney V6.1, produced images that looked like dolls or 3D renders, scoring as low as 3/10.
• Standout: MiniMax Image-01 and Recraft V3 managed to capture the softness of a child's features without crossing into the uncanny valley.

Best Models by Use Case

Based on the data, here are the recommended models for specific photorealistic scenarios:

📸 Best for Gritty & Detailed Portraits

• Recommendation: Midjourney v7 and Grok 2 Image
• Why: These models dominate when high-frequency detail (wrinkles, pores, tattoos) is required. In the Facial Tattoos prompt, Grok 2 Image achieved a 10/10 for "indistinguishable from a real photograph" textures.

💼 Best for Professional Headshots

• Recommendation: Recraft V3 and Nano Banana Pro
• Why: For the Businesswoman Headshot, Recraft V3 scored a perfect 10/10, nailing the depth of field and professional lighting without making the subject look artificial.

🌃 Best for Complex Lighting & Text

• Recommendation: ChatGPT 4o and Ideogram 3.0 (Quality)
• Why: When the prompt involves specific lighting (like split neon) or background text, these models excel. ChatGPT 4o scored 10/10 on the Neon Night Portrait for flawless split-lighting execution.

🌍 Best for Group Diversity

• Recommendation: Imagen 3.0 and Reve Image (Halfmoon)
• Why: In the Group Selfie, these models perfectly handled multiple subjects with distinct ethnicities and natural interactions, scoring 10/10 where others struggled with uniform "same-face" syndrome.

⚠️ Use with Caution

• DALL-E 3: While it adheres strictly to prompts, it consistently struggles with photorealism in this category, often producing a "plastic" or illustrative look.
• Flux 1.1 Pro Ultra: Excellent for general composition but prone to making young subjects look like dolls and failing text generation tasks.