Summary
Our comprehensive evaluation reveals a distinct hierarchy in current AI image generation, characterized by rapid advancements in prompt adherence, text rendering, and photorealism.
Top-Performing Models
- GPT Image 2.6 Sunburst: The undisputed champion with an overwhelming overall rating of 1439 Elo. It consistently dominates across almost every category and prompt type.
- GPT Image 2.5 Sunburst: The runner-up at 1325 Elo, maintaining excellent coherence and proving to be a highly capable predecessor to the 2.6 model.
- GPT Image 2: Secures third place with 1276 Elo, proving older architectures still hold substantial baseline power.
Major Trends
- The OpenAI Dominance: The top four spots are occupied by OpenAI models, showcasing a massive lead in side-by-side human preference.
- Text & Typography: The ability to render accurate text is now a major differentiator. Models that fail at text rendering plummet in rankings on complex graphic design tasks.
Surprising Discoveries
- Tight Mid-Tier Competition: Models like Grok Imagine 2.0 (1256 Elo), FLUX 3 Image (1235 Elo), and Muse Image (1229 Elo) are clustered closely together. Being within 30 points of each other indicates they are functionally tied in overall preference, often trading wins depending on the specific niche.
- High Performance from Alternate Providers: Qwen-Image-3.0-Pro (1218 Elo) and Nano Banana 2.1 (1215 Elo) put up a surprisingly strong fight, particularly in artistic and non-portrait tasks.
General Analysis & Useful Insights
By diving into the comparative data, several critical patterns emerge regarding what separates top-tier models from the rest of the pack.
Comparative Strengths
- Photorealistic Nuance: GPT Image 2.6 Sunburst thrives on realistic textures. For example, in the Elderly Woman Portrait, it achieved a near-perfect score (view image) for incredible photographic credibility, integrated reflections, and material detail.
- Artistic Emulation: The mid-tier models shine brightly in stylized tasks. Grok Imagine 2.0 and Nano Banana 2.1 demonstrate excellent color harmony and aesthetic flair when photorealism isn't the primary constraint.
Distinguishing Quality Factors
What makes a model score above 1300 Elo?
- Prompt Adherence: High-Elo models obey negative constraints and multiple complex instructions simultaneously.
- Coherence in Complexity: In Complex Scenes, lower-tier models resort to background blur or excessive fog to hide flaws, whereas top models maintain sharp, coherent details across crowded spaces.
- Anatomical Accuracy: A classic failure mode is the 'extra finger'. Models like Midjourney v7 often suffer severe penalties in Hands & Anatomy for hallucinating extra digits, severely tanking their win rates against safer models.
Common Failure Modes
- The Text Trap: In typography tasks, many models fail the basic spelling test. An image can be beautiful, but if a prompt asks for a specific phrase like 'Open 24/7' (Open 24/7 Sign) and the model hallucinates letters, it automatically loses to a less artistic but accurate model.
- Conceptual Reversals: In the Ultra Hard category, prompts like Astronaut ridden by a horse exposed models defaulting to their training biases (a human riding a horse) instead of following the absurd prompt. GPT Image 2.6 Flare succeeded brilliantly here (view image), proving its superior semantic understanding.
Best Model Analysis by Use Case
Different scenarios call for different tools. Here is a breakdown of model recommendations based on specific creative needs:
Photorealism & Portraiture
Category: Photorealistic People & Portraits
- Top Pick: GPT Image 2.6 Sunburst (1523 Elo). It commands a nearly 200-point lead over the runner-up, meaning it will win the vast majority of matchups. Its skin textures, lighting, and eye reflections are currently unmatched.
- Alternative: GPT Image 2 (1325 Elo) and Grok Imagine 2.0 (1303 Elo) are fantastic secondary options, offering highly polished, studio-quality lighting.
Typography & Graphic Design
Categories: Text in Images, Graphic Design
- Top Pick: GPT Image 2.5 Sunburst (1380 Elo in Text). Interestingly, the 2.5 version holds its ground exceptionally well here. It handles complex typography, neon signs, and branding lockups with incredible accuracy.
- Alternative: Qwen-Image-3.0-Pro (1311 Elo in Text) is surprisingly robust at embedding text cleanly into visual elements without obvious generative artifacts.
Artistic & Stylized Generation
Categories: Anime & Cartoon Style, Ghibli style
- Top Picks: GPT Image 2.6 Sunburst (1626 Elo in Ghibli) and MAI-Image-2.6 (1474 Elo in Ghibli).
- Analysis: If you need whimsical storytelling, vibrant 2D stylization, or atmospheric watercolors, this is where the competition gets interesting. While GPT Image 2.6 Sunburst still leads, models like MAI-Image-2.6 provide highly curated, gorgeous anime aesthetics that heavily appeal to human judges.
Complex & Ultra Hard Scenarios
Categories: Complex Scenes, Ultra Hard
- Top Pick: GPT Image 2.6 Sunburst (1738 Elo in Complex, 1524 Elo in Ultra Hard).
- Analysis: When requesting extreme edge cases, multi-subject interactions, or contradictory concepts, you must use the top-tier semantic engines. The massive 100+ point gaps here highlight that only the most advanced models possess the spatial coherence and logical reasoning required to handle heavy constraints without breaking down into surreal artifacts.
