Compare Image Editing Models for Optimal Choice

In recent weeks, nearly every major AI lab has released an image editing model. The first, FLUX.1 Kontext from Black Forest Labs, debuted in May and stood out for its style transformations and simple edits. Since then, we've seen a wave of new models, each strong in its own way. With so many options, it can be challenging to determine which model best suits your needs. In this article, we compare them and evaluate each across a range of image editing tasks.

Let's start with an overview of the cost and average inference time for each model. The cheapest is GPT-image-1 from OpenAI, starting at $0.01 per image, but it has the longest generation time of around 40 seconds. FLUX.1 Kontext [dev] (optimized by Pruna AI) is the fastest model at 1.9 seconds per generation, but hyper-optimized models come with a trade-off in image editing quality.

The first task we examined was object removal, a basic task that can be performed in Photoshop. We tested how different models performed when tasked with removing the Golden Gate Bridge from the image. The models SeedEdit 3.0 and Qwen Image Edit emerged as the winners, while FLUX.1 Kontext [pro] struggled with the task.

The next task involved changing the viewing angle of the object in the image. We aimed to get a front-facing view of the character and her cat. Only GPT-image-1 and Qwen Image Edit provided the desired view, although GPT-image-1 did not maintain character consistency.

Background editing requires models to understand object boundaries and generate coherent environments. The models SeedEdit 3.0 and Seedream 4 performed best, while Nano Banana showed the worst results, failing to maintain the character's integrity.

Text editing within images poses one of the most challenging tasks. We checked how models handled changing the word 'seven' to 'eight.' FLUX.1 Kontext [pro] and Nano Banana successfully integrated the word 'eight' naturally while preserving the original typography. Other models exhibited shortcomings in maintaining the original.

Style transfer showcases how each model understands artistic styles and applies them while preserving the content and composition of the original image. Some models excel at capturing fine artistic details, while others focus on maintaining structural integrity.