German start-up Black Forest Lab offers artificial intelligence models specialized in creating and editing images for professionals. FLUX.1 Kontext (max), the latest model, perfectly illustrates this strategy.
If text generation models still remain largely the prerogative of American companies, in the field of images, the start-up Black Forest Lab is doing well. Founded in 2024 in eastern Germany, a few kilometers from Colmar, by former employees of Stability AI, the company specializes in the creation of foundation models dedicated to the creation and editing of images. From 2024, the start-up is launching its FLUX family, with immediate success.
The models are so qualitative (adhesion to the prompt, visual consistency, etc.) that they will end up attracting xAI and Mistral: FLUX models were used by default by Grok and on Le Chat de Mistral (still today) to generate images. But it is certainly in the creative community that Black Forest Lab models have the best reputation.
FLUX.1 Kontext, the latest
The latest model, FLUX.1 Kontext offers cutting-edge performance, whether in raw generation or image editing. The model is particularly relevant for accurately following initial instructions (prompt adhesion), generating visually coherent scenes and producing near-perfect photorealistic content. It is also quite good at generating typographic content in images (a difficulty still persisting in text-to-image models). Finally, FLUX.1 Kontext offers one of the fastest generations on the market with Gemini Nano Banana, in our experience.
FLUX.1 Kontext currently exists in three different versions:
- FLUX.1 Kontext (pro), the standard model which allows you to create and edit visuals very quickly.
- FLUX.1 Kontext (max), which allows you to create and edit images with greater precision than (pro) but taking a little more time. It is this model that offers the best visual results.
- FLUX.1 Kontext (dev), model of 12 billion parameters, which offers capabilities close to (pro) in open source version. It will therefore be slightly less optimized than (pro) and even more than (max)
The great strength of the FLUX.1 Kontext family lies in the available ratios. It is possible to generate images in 1:1, 16:9, 4:3, 3:2, 9:16, 3:4, 2:3, 21:9… A real asset for the production of professional content.
Image generation, the JDN test
As part of our test, we choose to measure the performance of FLUX.1 Kontext (max), the SOTA model of the family and the most used, in the top 10 of the best editing and image generation models of Artificial Analysis for months. We compare it to the best proprietary models of the moment, namely Imagen 4 from Google for image generation and Nano Banana for the editing side.
1.A photorealistic image of a race on a ski slope
Prompt: an epic car race taking place on a snowy ski slope. Rally cars speed downhill, kicking up clouds of snow that splash over the cheering crowd. Spectators stand behind safety barriers. The scene is electric a mix of bright white snow, and the roar of engines echoing through the mountains.
On pure visual consistency and adherence to prompt, Flux and Imagen are tied. Both models offer images that perfectly respect the instructions. On interpretation and the pure photorealistic aspect, Flux wins. Black Forest Lab’s AI offers a more coherent scene and a fairly relevant angle, similar to what a real photographer could have done.
2.A portrait, oil on canvas, of Henry IV
Prompt: a historical oil painting of King Henry IV of France on his majestic white horse. The king, dressed in regal attire, is being served a glass of wine by an elegant servant standing beside the horse. The servant holds the bottle with a white cloth, its label reading ‘Cheval Blanc’. The scene is rendered in the classical 17th-century style, with warm lighting, detailed brushwork, and rich, textured colors.
On a photorealistic level, Flux offers a more realistic and credible version of a real oil on canvas. On the other hand, in terms of typography, only Imagen manages to produce the text “Cheval Blanc” on the wine bottle. So point goes to Google for accuracy.
3.A vintage movie poster
Prompt: a vintage 1980s movie poster titled “Before AI”. The artwork shows a group of programmers working intently at bulky CRT computers, surrounded by cables, floppy disks, and old IBM machines. The scene captures the nostalgic atmosphere of retro sci-fi films, with neon lighting, blue tones, and a grainy film texture.
Here again, Flux seems to have a slight lead in overall realism. The grain of the poster generated by the German model is more worked. Imagen offers an image with strong contrasts. Flux narrowly wins on overall realism. However, both models produce unreadable text at the bottom of the page.
Can Flux beat Nano Banana at image editing?
Published well before the Google model, can FLUX.1 Kontext (max) compete with the latest from Google? This is what we will try to find out.
1.Edit an element on an image
prompt: “the Mona Lisa wearing an elegant pink hat, seamlessly integrated into the original painting’s style. The hat should look authentic, oil-painted with soft tones and realistic texture, matching Leonardo da Vinci’s lighting and color palette”
For this first editing task, Nano Banana offers the most realistic version. Google’s model manages to add a pink hat to Mona Lisa without changing the nature of the painting. The facial features are generally well preserved and the style well respected. Flux offers a more modern version but modifies the original image quite profoundly. Damage.
2.Replace a background
For our final editing test, we chose to replace the lunar background of Buzz Aldrin’s famous photo with a Martian setting.
Prompt: replace the lunar background in this image with a Martian landscape, and update the reflection on the astronaut’s helmet to match the new Mars environment.
Here again, the advantage goes to Nano Banana, which manages to replace the background with remarkable realism, without altering the slightest detail of Buzz Aldrin’s suit.
Flux nevertheless marks a positive point: the light, slightly reworked, appears more coherent in its version.
FLUX.1 Kontext (max), an excellent template for image creation
FLUX.1 Kontext (max) stands out as an excellent model for creating images. If you have to choose between several solutions to generate precise and photorealistic visuals, you can easily turn to it. For developers, the model is accessible via API at a price of $0.08 per image, compared to $0.06 for Imagen 4 Ultra. A difference in price which is amply justified given the quality of the renderings. The template is also recently available in Adobe Photoshop.
On the other hand, FLUX.1 Kontext (max) performs less well in image editing. In this area, Nano Banana remains the undisputed leader for the moment. Google’s model manages to modify elements while preserving the integrity and style of the original image, where Flux still struggles to maintain overall consistency. For retouching existing images, it is therefore better to turn to Google’s solution.
