Description
The artwork presented here is an interesting example of the state of fast AI image generation in the first month of 2025.
I used the FLUX Schnell generative text-to-image model because I believe it is one of the most economical solutions available today. However, while generative models like FLUX Schnell can create usable images, the objects they can generate are ultimately limited by what they have learned from their training data.
The outputs of the models are constrained by the concepts present in the training data. If an object or a concept is absent or poorly represented, the model may render it inaccurately, or not at all.
Sea buckthorns are not typically included in the training data for image models, or at least there may not be many images featuring such plants in the model's training set. This is likely true for all the models I tried, as none of them was able to render proper common sea buckthorns.
On the contrary, ask a model to render roses, for example, and you'll receive beautiful roses, even when using older stable diffusion models. However, show me a model that excels at rendering sea buckthorns in 2025.
final artwork
base image
glass of the bottle
In some generated images, the orange berries were somewhat close to the real ones, but the leaves were never exactly the same as those of the actual plant.
FLUX Schnell could only generate a limited number of images with usable common sea buckthorn berries, and even those berries had to be photo-retouched to resemble real ones. The image I chose as the base for the artwork had other issues as well, so there was a lot to fix. Additionally, the glass bottle in the main image was quite off, but luckily, its cap was fairly close to the real one, so I could keep that.
I had to generate an image that featured a glass with the correct dimensional proportions. After obtaining such an image, I merged the two in my image editing program.
All that was left was to extend the square image and fix any remaining issues. To extend the image, I used some other similarly generated images that were otherwise unusable, but from which I could manually extract certian parts.
Lastly, I manually applied the label to the bottle. The label was rendered in Blender 3D.
Overall, it was not straightforward to use generative text-to-image models to create this artwork, but with considerable effort, I managed to produce a visually pleasing, photo-like picture, I think.