Did ChatGPT Just Kill Nano Banana?

Quick Overview

ChatGPT Image 1.5 generally outperforms Nano Banana in image editing tasks, correctly executing complex instructions like removing subjects while preserving pose, lighting, and background, although Nano Banana sometimes maintains better visual consistency across iterative edits, as demonstrated in tests involving multi-step edits and style transfers.

Key Points: ChatGPT Image 1.5 successfully executed complex editing prompts, such as removing a person while maintaining pose, lighting, and background integrity (09:10). In a family dinner photo edit, ChatGPT aged the adults by 15 years while successfully preserving the snow visible outside the window (11:16). When generating a multi-era portrait from a single input image, ChatGPT accurately applied three distinct styles (Renaissance, 1980s studio, Modern cinematic) while maintaining facial structure (19:21). Nano Banana occasionally struggled with complex spatial instructions, such as generating a 3x3 grid where it produced a 4x4 grid (15:34). Nano Banana also struggled with maintaining brand consistency across three images (storefront, cup, coffee bag) for a fictional brand named 'EMBER' (21:03). The speaker notes that ChatGPT's ability to maintain context through multi-step edits is a significant improvement over previous models (15:08).

Context: The video reviews and compares the image generation and editing capabilities of the newly released ChatGPT Image 1.5 model against its predecessor, Nano Banana (likely referring to a competitor or previous iteration). The presenter, wearing a 'Future Tools' hat, tests the models using complex prompts involving subject removal, iterative editing, style transfer, and adherence to specific layout/text requirements to gauge improvements in instruction following and consistency.

Detailed Analysis

The video concludes that ChatGPT Image 1.5 shows significant advancements, particularly in image editing and following complex, multi-step instructions, often outperforming Nano Banana in these areas. When testing subject removal and modification on a photo of three men (08:57), ChatGPT correctly removed the specified person, changed the clothing, and added neon lighting while maintaining the original pose and background, although it slightly altered the pose. In another test involving aging a family photo by 15 years, ChatGPT successfully aged the adults while keeping the winter scene outside the window (11:16), although it failed to age the young girl. A complex layout prompt asking for a 3x3 grid of specific items resulted in Gemini generating a 4x4 grid (15:34), whereas ChatGPT followed the 3x3 layout correctly (15:21). In brand consistency tests, ChatGPT successfully applied the same minimalist black/orange branding across a storefront sign, coffee cup, and coffee bag (21:03), while Nano Banana struggled with consistency across the three items. The speaker notes that ChatGPT's ability to maintain context across sequential edits is a major advantage over previous models.

Raw markdown version of this recap