Recraft V3

Summary for Recraft V3

Recraft V3 currently holds an overall Elo rating of 875, placing it 29th out of 40 models on the leaderboard. It sits in the lower-middle tier, significantly behind the 1300+ Elo leaders like GPT Image 2.5 Sunburst.

Key Findings & Trends:

• Beautiful but Disobedient: The model is a master of surface-level aesthetics, textures, and lighting, but struggles significantly with prompt adherence and complex conceptual logic.
• Surprising Highs: Despite a very low category Elo (677) in Photorealistic People & Portraits, it achieved stunning 1st place finishes in specific tight portraits, such as Facial Tattoos and Freckled B&W.
• Style Stubbornness: It completely ignores requests for specific 2D art styles (like Studio Ghibli or pixel art), defaulting instead to hyper-realism, stylized 3D, or generic storybook illustrations.

Quick Conclusion: Use Recraft V3 for close-up, high-fidelity portraiture and simple graphical textures, but avoid it for complex compositions, strict 2D style emulation, or logic-defying surrealism.

General Analysis & Useful Insights

Comparative Strengths

Recraft V3 is incredibly strong at rendering high-fidelity materials, lighting, and textures. When examining the detail execution scores across prompts, the model frequently earns 8s and 9s. It renders skin pores, fabric weaves, and dramatic cinematic lighting beautifully. Its relative best performances on the leaderboard are in Text in Images (988 Elo) and Hands & Anatomy (950 Elo), showing it can handle structured, physical details reasonably well compared to its baseline.

Quality Factors

The model distinguishes itself through a photographic finish. In the Businesswoman prompt, it delivered polished, aspirational lighting that earned a top-3 spot. Its native aesthetic leans heavily toward a clean, commercial, and highly rendered look, making it excellent when the prompt aligns with this default style.

Common Failure Modes

1. Conceptual Reversal: The model struggles with non-standard relationships. For example, in the Astronaut Ridden by Horse prompt, it generated a standard astronaut riding a horse, completely missing the requested absurdity.
2. Literal vs. Surreal: When asked for a Galaxy Waterfall that pours stars instead of water, it just drew a beautiful normal waterfall under a starry sky.
3. Anatomical/Counting Errors: Complex intersections break down easily. It generated three hands in the Heart Shape prompt, resulting in a critical failure.
4. Style Erasure: It fails to emulate distinct retro styles like SimCity Pixel Art or Apple II, delivering smooth vectors or generic modern desktops instead.

Best Model Analysis by Use Case

Where Recraft V3 Excels

• Tight Photorealistic Portraits: While it struggles with wide or complex human scenes, it is highly recommended for close-up character work. It handles intense lighting and skin details masterfully, easily competing with top models on prompts like Freckled B&W.
• Basic Graphic Elements: In the Graphic Design category, it performs adequately for simple patterns and textures. For example, it scored an 8 on the Art Deco Pattern prompt, providing a clean and highly usable seamless texture.
• Simple Typography: With a 988 Elo in Text in Images, it is competent at straightforward neon signs or billboards, rendering words sharply when they don't require complex physical integration.

Where to Avoid

• Anime & Cartoon Style & Ghibli Style: The model fundamentally refuses to generate authentic 2D cel-shaded animation styles. Whether asked for Kiki's Delivery Service or a 90s Space Battle, it outputs highly rendered 3D models or textured storybook illustrations. It holds a dismal 641 Elo in Anime.
• Complex Scenes: With a low 695 Elo here, the model falls apart when asked to juggle multiple subjects with distinct actions. In the Classroom and Beach prompts, it merged details into ambiguous shapes and entirely missed the requested simultaneous activities.
• Logic Puzzles & Ultra Hard: The model fails to synthesize complex multi-step instructions, usually ignoring the hardest constraints (like making vines form actual letters in the GROWTH prompt) to focus merely on making the image look aesthetically pleasing.