ai nano banana 2.1 google gemini image generation benchmarking diffusion models text rendering computer vision generative ai technical analysis

Benchmarking Google’s Nano Banana 2.1: Evaluating Instruction Following, Textual Rendering, and Identity Preservation against GPT Image 2.5

5 min read

Benchmarking Google’s Nano Banana 2.1: Evaluating Instruction Following, Textual Rendering, and Identity Preservation against GPT Image 2.5

The release of Google's Nano Banana 2.1 marks a significant shift in the competitive landscape of generative image models. According to official documentation, the 2.1 iteration is engineered to outperform all previous iterations—including Nano Banana Pro and Nano Banana 2—while operating at a cost-efficiency four times greater than its predecessors. The primary technical claims center on three pillars: enhanced instruction following (prompt adherence), superior in-image text rendering, and improved character consistency through grounding in Google Image Search datasets.

To validate these claims, we conducted a multi-model comparative analysis using the OpenArt platform, pitting Nano Banana 2.1 against Nano Banana Pro, Nano Banana 2, and OpenAI’s GPT Image 2.5 across several high-complexity benchmarks: semantic density, spatial reasoning, physical motion simulation, and identity preservation in image-to-image workflows.

Benchmark 1: Semantic Density and Textual Rendering Accuracy

The first test involved a highly complex prompt requiring the generation of a photorealistic, street-level scene at a trendy coffee shop. The objective was to evaluate the model's ability to render legible, non-repetitive text across multiple planes (signage, windows, and handheld objects) while maintaining environmental consistency in an overcast/rainy setting.

The results were mixed for the new 2.1 architecture. While GPT Image 2.5 demonstrated high legibility, it struggled with semantic repetition—specifically repeating "good coffee, good people" across various window panes. Nano Banana 2.1 showed significant improvements in text clarity on small-scale objects like coffee cups; however, it exhibited a tendency toward "overly neat" rendering, where the typography lacked the organic imperfections found in real-world photography. Notably, some models (specifically Nano Banana 2) suffered from catastrophic failure in text rendering, where characters became illegible or structurally collapsed.

Benchmark 2: Spatial Geometry and Low-Angle Perspective Precision

A critical metric for advanced diffusion models is spatial reasoning—the ability to interpret precise geometric instructions regarding camera placement and object orientation. The prompt required a specific low-angle shot (30cm above pavement, 20-degree tilt) featuring a woman holding a red umbrella in her left hand and a paper cup in her right, with a white sedan positioned horizontally behind her.

This benchmark revealed the current ceiling of generative capabilities. While Nano Banana Pro achieved the highest degree of cinematic atmosphere and relatively accurate camera positioning, all models struggled with complex anatomical and spatial constraints. Specifically:

  • Hand Anatomy: Multiple models failed to correctly assign objects to the specified hands (e.g., placing the umbrella in the right hand instead of the left).
  • Spatial Orientation: The distinction between driver-side and passenger-side orientation for the sedan was frequently conflated.
  • Depth Perception: Maintaining the correct proximity of foreground elements, such as a gas pump near the lens, proved difficult for Nano Banana 2.1 and GPT Image 2.5.

Benchmark 3: Dynamic Motion and Physical Simulation

To test the models' ability to simulate fluid dynamics and kinetic energy, we prompted a hyper-realistic photograph of an individual jumping on a backyard trampoline. This required the model to handle complex physics: hair movement, clothing deformation (a loose gray t-shirt), and the trajectory of secondary objects (a water bottle).

Nano Banana 2.1 demonstrated superior performance in simulating motion blur and kinetic realism, particularly regarding the upward flow of hair and the realistic "riding up" effect of the fabric during the jump. In contrast, GPT Image 2.5 and Nano Banana Pro struggled with structural integrity; for instance, one iteration failed to render the trampoline safety net entirely, while another exhibited pixelated artifacts in the background foliage (tree textures), bordering on a low-fidelity pixel art aesthetic.

Benchmark 4: Information Retrieval and Infographic Synthesis

Moving from purely aesthetic generation to functional data visualization, we tasked the models with generating an infographic regarding decompression sickness ("the bends"). This test evaluated "zero-shot" research capabilities—how well a model can retrieve factual information and organize it into a legible, visually coherent layout without explicit stylistic prompting.

In this category, Nano Banana 2.1 emerged as a leader in both legibility and aesthetic design. While Nano Banana Pro produced accurate but "boring" (low-complexity) layouts, and GPT Image 2.5 suffered from significant text garbling (character substitution errors like 'i' for 'l'), Nano Banana 2.1 maintained high typographic integrity alongside sophisticated chart integration. This suggests that the 2.1 update has significantly bolstered its ability to handle structured, information-dense prompts.

Benchmark 5: Identity Preservation in Image-to-Image Workflows

The final benchmark focused on character consistency and identity preservation using a reference image of the creator. The task involved two distinct modes: generating a stylized thumbnail (a boxer banana punching the subject) and complex image editing (changing clothing, location to Tokyo at sunset, and adding accessories).

In the thumbnail generation, Nano Banana 2.1 outperformed GPT Image 2.5 in facial fidelity. The latter often applied an "uncanny" smoothing effect that drifted away from the source identity. In the more complex editing task—which required changing a shirt to a flannel/t-shirt layer and altering the entire environmental lighting to sunset—Nano Banana 2.1 and Nano Banana 2 were the most successful at maintaining the subject's facial features while seamlessly integrating new lighting conditions and clothing textures.

Conclusion

The data suggests that Google’s Nano Banana 2.1 is a significant step forward in terms of cost-efficiency and information density (infographics/text). While it has not yet solved the fundamental challenges of complex spatial reasoning and precise anatomical manipulation, its ability to maintain identity preservation and handle high-complexity text rendering makes it a formidable competitor to GPT Image 2.5. For developers looking for a balance between high-fidelity output and low inference cost, Nano Banana 2.1 represents a highly optimized solution.