NEW GPT Image 1.5 vs Nano Banana Pro

GPT Image 1.5 vs. Nano Banana Pro: A Detailed Comparison

Key Concepts:

• GPT Image 1.5: OpenAI’s latest AI image generator, integrated with ChatGPT, offering image generation, editing, and manipulation capabilities. Free to use with usage limits.
• Nano Banana Pro: Currently considered a leading AI image generator, known for its high-quality outputs and advanced editing features.
• Deepfake: The creation of realistic but fabricated images or videos, often involving swapping faces.
• Outpainting: Expanding an existing image beyond its original boundaries.
• Aspect Ratio: The proportional relationship between an image's width and height.
• Segmentation Map: An image where different objects or regions are identified and labeled with distinct colors.
• Depth Map: An image representing the distance of objects from the camera, often used for 3D effects.
• Guardrails: Safety mechanisms implemented in AI models to prevent the generation of harmful or inappropriate content.
• Sprite Sheet: A collection of images used in animation, often representing different frames of a character's movement.

I. Introduction & Capabilities

OpenAI has released GPT Image 1.5, a new image generator directly integrated with ChatGPT. The tool allows for image editing (hairstyle, clothing changes), background alterations, perspective shifts, deepfake creation, and stylistic adjustments. The video focuses on a rigorous comparison between GPT Image 1.5 and Nano Banana Pro, currently considered the leading AI image generator. The presenter aims to test the limits of GPT Image 1.5 beyond basic editing tasks, focusing on complex prompts. The tool is available for free within ChatGPT, with usage limits on the free plan (approximately a dozen images per day).

II. Pokémon Identification & Accuracy

The first test involved generating a 4x4 grid of Pokémon based on provided Pokédex numbers. Both GPT Image 1.5 and Nano Banana Pro successfully identified most Pokémon. However, Nano Banana Pro rendered the images with slightly better quality, particularly for Pokémon like Golbat and Pidgeotto. Discrepancies arose with Pokémon #2011 (Unknown) and #301 (Electabuzz), where both models struggled with accurate representation. Nano Banana Pro demonstrated greater accuracy in these cases, while GPT Image 1.5 sometimes fabricated details.

III. Emotional Expression Rendering

A more challenging test involved generating a 4x4 grid depicting a young woman expressing 16 different emotions (happiness, awe, enthusiasm, etc.). GPT Image 1.5 performed remarkably well, accurately conveying even subtle emotions like relief, anticipation, and nostalgia. Nano Banana Pro struggled with some of the more nuanced expressions (confidence, awe, nostalgia, pride, anticipation), leading the presenter to award the point to GPT Image 1.5.

IV. Solving Math & Biology Assignments

GPT Image 1.5 demonstrated an unexpected ability to solve math problems presented in an image format, even replicating messy handwriting to appear as if the work was done manually. Nano Banana Pro also succeeded, but GPT Image 1.5’s generation didn’t alter the background of the original assignment. However, both models failed when presented with a biology worksheet requiring the labeling of cell organelles, producing inaccurate and nonsensical results.

V. Advanced Image Manipulation: Quadrants & Inversion

The presenter tested the models’ ability to divide an image into four quadrants: infrared thermal map, segmentation map, depth map, and a color-inverted version. Nano Banana Pro consistently outperformed GPT Image 1.5. While GPT Image 1.5 produced a reasonable infrared thermal map, its segmentation and depth maps were less accurate, and the color inversion was imperfect. Nano Banana Pro’s results were more precise and visually accurate across all quadrants.

VI. Clock & Wine Glass Challenge

A difficult prompt requiring a clock displaying 11:15 and a wine glass filled to the top proved challenging for both models. GPT Image 1.5 successfully generated both elements, though the hour hand was slightly off in multiple attempts. Nano Banana Pro correctly rendered the wine glass but struggled with the precise time on the clock. This test resulted in a tie.

VII. Rare Frog Identification & Accuracy

The presenter tasked the models with generating images of the four rarest frog species, including their scientific names and descriptions. Both models failed significantly. GPT Image 1.5 provided inaccurate information and generated images that didn’t match the actual species. Nano Banana Pro fared slightly better, correctly identifying some species visually but misclassifying others as “rare” when they were not.

VIII. Manga Generation & Translation

Both GPT Image 1.5 and Nano Banana Pro successfully generated a coherent, multi-panel manga page based on a given prompt. However, when tasked with colorizing the manga and translating it to Chinese, GPT Image 1.5 produced errors in character preservation and inaccurate Chinese translation. Nano Banana Pro provided a more accurate translation and better preserved the original manga style.

IX. Interface Generation: YouTube Search Results

GPT Image 1.5 excelled at generating a realistic screenshot of YouTube search results for "cute cats," with minimal errors. Nano Banana Pro’s generation contained numerous misspellings and inaccuracies. This test awarded the point to GPT Image 1.5.

X. Generating Existing People & Character Sprite Sheets

GPT Image 1.5 encountered “guardrails” and refused to generate a group photo of numerous celebrities. Nano Banana Pro successfully generated the image. When prompted to generate images of the top 16 richest people, GPT Image 1.5 struggled with accurate facial representations, while Nano Banana Pro produced more recognizable results. GPT Image 1.5 can generate transparent images (PNG files) for sprite sheets, a feature Nano Banana Pro lacks.

XI. Anime Character Generation & Final Fantasy 9 Remaster

Both models successfully generated a group of anime characters (Amelia, Gojo, Nezuko, etc.). Nano Banana Pro demonstrated slightly better character consistency, accurately depicting details like Bart Simpson’s four fingers. When tasked with creating a “faithful remaster” of a Final Fantasy 9 scene, both models performed well, but Nano Banana Pro’s output was considered slightly more polished.

XII. Table to Chart Conversion

GPT Image 1.5 failed to accurately convert a complex table into a chart, omitting data and producing incorrect values. Nano Banana Pro successfully converted the table, performing calculations (e.g., percentage calculations) and creating a visually accurate chart. This was a clear win for Nano Banana Pro.

XIII. Wheatstone Bridge Circuit & Where's Waldo

Both models struggled to accurately generate a diagram of a Wheatstone bridge circuit. Both also failed to create a high-resolution, complex "Where's Waldo" image, producing distorted details and easily identifiable characters.

XIV. Specifications & Conclusion

GPT Image 1.5 is free to use within ChatGPT, with usage limits on the free plan. It generates images up to 1.5K resolution and currently supports limited aspect ratios. While GPT Image 1.5 represents a significant improvement over previous versions, eliminating the yellow tinge and demonstrating strong prompt understanding, Nano Banana Pro remains the superior AI image generator based on the presenter’s testing. Independent leaderboards (Artificial Analysis, LM Arena) currently rank GPT Image 1.5 as #1, but the presenter believes this ranking may shift as sample sizes increase. The presenter encourages viewers to experiment with GPT Image 1.5 and share their results.

Notable Quotes:

• “I really want to test its limits. So, I'm going to feed it some really tricky prompts that even trip up the best image editors.”
• “At least from my examples, I would have to say Nano Banana Pro is still the king of AI image.”
• “They got rid of the yellow tinge, which is awesome.”

Synthesis:

GPT Image 1.5 is a powerful and accessible AI image generator, offering a compelling free alternative to existing tools. While it excels in certain areas (emotional expression, solving simple math problems, interface generation), Nano Banana Pro consistently demonstrates superior performance in complex tasks requiring accuracy, detail, and world knowledge. The presenter acknowledges the potential for GPT Image 1.5 to improve with further development and data, but currently, Nano Banana Pro remains the benchmark for AI image generation quality.