FLUX.2 (klein) offers an on-device alternative to Gemini’s Nano Banana Pro for image editing.
This is another performance from Black Forest Labs. The German start-up specializing in generative AI applied to images unveiled a new model in mid-January capable, on paper, of competing with American image editing models (Nano Banana Pro or gpt-image-1.5 from OpenAI). His strength? The template can be used locally on your machine, without requiring you to send your images to the cloud. Two versions are offered: an Apache 2.0 version (free for any use) and a version without commercial use (open weight).
A version 4B and 9B, two licenses
Black Forest Labs is therefore deploying two distinct models to cover all needs: a compact version of 4 billion parameters and a premium version of 9 billion parameters. Version 4B offers total freedom of use, including commercial use, without any particular restrictions. The model requires approximately 8.4GB of VRAM for the quantized model (9.2GB for the base version). Its main advantage lies in its particularly fast inference time. Black Forest talks about 0.3 seconds on Nvidia GB200 and 1.2 seconds on RTX 5090 for the distilled version. Its relatively modest size for an image model, however, restricts image quality. The model nevertheless excels in use cases requiring minimal latency, for minimal edits in the images.
For its part, the 9B version represents the flagship of the range. It is capable of competing with models five times larger. Even though the weights are available as open source, on Hugging Face, use of the model is restricted by the “Non-Commercial License.” The latter exclusively authorizes non-commercial uses, research, testing or charitable purposes. For any commercial exploitation, the acquisition of a paid license becomes obligatory. The license starts at $199 for 10,000 images (or $0.02 per image). For comparison, Gemini charges its image generation API $0.134 per image at 1K/2K resolution and $0.24 at 4K, or approximately $1,340 to $2,400 per 10,000 images generated. On the hardware side, the model requires approximately 19.6GB of VRAM for the quantized version (21.7GB for the base).
The JDN test
To evaluate the real performance of FLUX.2 (klein), we chose to use the unquantized version 9B, the most efficient in the Black Forest Labs catalog. The objective is to compare the model to the current leader in image editing, Nano Banana Pro from Gemini, on four concrete use cases.
1. Change the color of an element: the point for Klein
For this first use case, we will start in easy mode by simply changing the color of an element present in an image. Here we are going to modify the color of this Rafale, to go from gray to an amaranth red.
Prompt: Changes the color of the plane to amaranth red while exactly preserving the bright reflections of the sun on the wings and fuselage.
FLUX.2 (klein) clearly wins this first test. The model manages to apply amaranth red while preserving the original light reflections, particularly on the cockpit and the upper part of the fuselage. Conversely, Nano Banana Pro standardizes the color too much and crushes the details, losing (a little) the realism of the scene.
2. Replace a character: Nano Banana in front
More complex, we ask the AI to replace a human in an image, all cleanly. For this example, we start from a classic photograph: a couple walking in a clearing. The goal is to replace the woman with a ballet dancer in full attire – white tutu, tight bun, pointe shoes – while retaining the essence of the original scene.
Prompt: replaces the woman on the left with a classical ballet dancer in full outfit: white tutu, tight bun, pink pointe shoes. Maintains exact position with the man, body angle and torque dynamics.
The image generated by FLUX.2 klein certainly creates a spectacular composition, but completely destroys the initial dynamic of the couple walking together. Nano Banana Pro maintains the basic scene, while integrating ballet elements (visible tutu, posture). The point is therefore awarded to Nano Banana.
3. Add an external element: Nano Banana still in front
The exercise gets tougher, it now involves introducing a completely foreign object into a scene, while maintaining perfect visual coherence. For this demonstration, we start from an American desert road stretching as far as the eye can see and add a concert grand piano.
Prompt: Adds a shiny black grand piano (concert type) placed in the middle of the deserted road, about 50 meters from the camera. The lid is raised, the keyboard visible. Naturally incorporates this surreal element while retaining the perspective of the road, the white marking lines, the reddish mountains on the horizon, the cloudless sky, and the atmosphere.
FLUX.2 (klein) completely reinterpreted the scene by modifying the sky, the lighting atmosphere and accentuating the red tones of the mountains. Nano Banana Pro adds the piano while preserving the original orange sky and sunset ambiance. The only positive point for Klein is that the integration of the piano is technically superior: the reflections on the instrument, the cast shadow and the visual anchoring on the road are more convincing.
4. Day/night transformation: Klein returns
For this last test, we ask the AI to transform a daytime photograph into a nighttime photograph. In this case, going from Stonehenge with full day light to subdued night light with a full moon.
Prompt: Transforms this daytime scene into a full moon night. Replaces the blue sky with a deep night sky with a bright full moon positioned above the stone circle.
klein is photographically more realistic. On a backlit night with a full moon, Stonehenge would most certainly appear in silhouette like this. This is exactly what you would get in a real photo. Nano Banana Pro artificially keeps too much detail and texture on the stones when they should be in shadow. Klein respects the physics of night light, even if this sacrifices the details of the original scene.
When to use FLUX.2 (klein)?
Nano Banana Pro maintains a good lead over FLUX.2 (klein), particularly in its ability to preserve the exact coherence of the original scene, a criterion which remains fundamental in image editing. In our four tests, Google’s model systematically preserves the structure, atmosphere and dynamics of the source image, where Klein has an unfortunate tendency to reinterpret the scene in his own way. Concretely, Nano Banana Pro remains the choice to make when you need surgical precision on the initial composition, even if the final rendering can sometimes lack finesse. Klein, on the other hand, produces visually more accomplished results in terms of realism and textures, but at the cost of less fidelity to the original image. It will then be necessary to agree to multiply the iterations to obtain a result which combines visual quality and respect for the initial framework.
But the real strength of FLUX.2 (klein) lies in another area. Firstly, on the price side, the model stands out as a significantly more economical option, particularly suited to batch processing. But the key argument is undoubtedly that of data sovereignty. Unlike Nano Banana Pro which requires you to go through Google’s American cloud servers, FLUX.2 (klein) runs entirely locally on your own infrastructure. No images are sent externally. A considerable advantage in the age of digital sovereignty.
