The burgeoning field of AI image generation witnessed a compelling showdown as Matthew Berman put Qwen's new open-source model, Qwen-Image-Edit, through a rigorous battery of tests against established players like Nano Banana, GPT Image 1, and Seedream 4.0. Berman's comprehensive comparison, covering a diverse array of image manipulation tasks from scene composition to intricate material transformations and stylistic rendering, unequivocally demonstrated that the current AI landscape boasts specialized strengths rather than a single, dominant champion.
The testing methodology involved presenting each model with an initial image or two, coupled with a detailed prompt outlining the desired transformation. For instance, the first challenge involved compositing a portrait into a waterfall setting, demanding natural lighting and mist effects. While Qwen-Image-Edit produced a "pretty good" result, it was GPT Image 1 that truly shone, with Berman noting, "This is by far the best version. It looks the least real in my opinion, but in terms of style and consistency across the image and into the background, it is the best." This early observation underscored a recurring theme: the subjective nature of "best" in AI image generation, often hinging on specific stylistic preferences or technical requirements.
A core insight emerging from Berman's analysis is the absence of a universal winner. Each model exhibited distinct aptitudes. Qwen-Image-Edit, for instance, proved remarkably adept at integrating objects into new environments with believable lighting and sand displacement, as seen in its superior performance transforming an SUV into a desert scene. Berman exclaimed, "Qwen Image Edit Plus. This looks phenomenal. Absolutely amazing," highlighting its ability to convincingly blend elements and simulate environmental effects like sun-blasted light and sand. Conversely, Nano Banana struggled with this task, appearing as if the SUV was "just plopped in the desert."
However, Nano Banana found its stride in tasks demanding meticulous consistency and recoloring, such as changing a car's paint to metallic midnight blue while preserving chrome trim and reflections. It also excelled at recoloring kitchen cabinets to sage green, accurately maintaining the original wood grain texture. In contrast, Qwen-Image-Edit, while good, sometimes lost the underlying texture. This showcases the specialized strengths different models possess.
The challenge of placing an executive headshot into a modern office further illuminated these distinctions. Qwen-Image-Edit adeptly integrated the individual, realistically positioning them in a chair with accurate lighting. However, competitors like Nano Banana, Seedream 4, and GPT Image 1 largely failed, merely superimposing the face onto the background without regard for perspective or interaction. This highlights a crucial divergence in their underlying architectural capabilities for understanding 3D space and object interaction. Berman remarked on the overall variability, stating, "It's crazy to see how different some of these models perform depending on the prompt."
Related Reading
- Meta's AI Talent Raid Exposes Apple's Critical Juncture in Generative AI
- Surfing the AI Tsunami: Strategic Navigation for Disruption
Beyond these primary insights, the tests revealed common pitfalls. Many models struggled with precise numerical instructions (e.g., 2.3cm above the wrist) or maintaining facial texture during stylistic transformations. The "bleeding edge" tests, designed to push the models to their limits, frequently resulted in humorous failures, such as Nano Banana's attempt at placing puppies on a beach, which Berman dismissed as a "complete fail" that "looks like somebody who is using Photoshop for the first time made this."
Ultimately, Berman's experiment serves as a vital guide for founders, VCs, and AI professionals navigating the complex landscape of generative AI. It underscores that relying on a single model for all tasks is often inefficient. Instead, the pragmatic approach involves leveraging specialized tools for specific needs. The video concludes by advocating for experimentation, offering an open-source script that allows users to replicate these tests and identify which model best suits their unique image generation and editing requirements. The era of one-size-fits-all AI is rapidly giving way to a mosaic of specialized, powerful solutions.
