This is our open-source "Banana," a small Modle capable of contextual reasoning.
It is recommended that the input image resolution be at least 1920X1080,
as larger-sized input images can effectively improve reasoning quality.
The reasoning speed is not particularly fast, and the application supports output of any aspect ratio up to 2K.
However, 1280X720 remains the most efficient resolution.
The greater the weight of the main subject, the stronger the "imagination" (of the application).
But an excessively high weight may cause image breakdown or deviation of character features. The default value is 0.4.
It is possible to mount a style LoRA (Low-Rank Adaptation) or the contextual LoRA from the previous version.
Nevertheless, if the weight is too high, it will lead to the deviation of character features. The default value is 0.3.
Prompt: Advertising design – This can be used for advertising design.
Prompt: Shot/reverse shot – This can be used for shot/reverse shot (a filming technique).
Prompt: Another angle – This function can still fix the scene and find other visual angles, making it highly practical.
Once this prompt is entered, the "camera" (in the application) will not be able to leave the current scene.
Prompt: Hand close-up – This is used to switch to a close-up shot of hands.
Prompt: Imagine – This can improve reasoning quality to a certain extent, but limited by the scale of training samples,
the actual improvement is not as significant as expected!
Prompt: OSM style – This can improve reasoning quality to a certain extent, but limited by the scale of training samples,
the actual improvement is not as significant as expected!
The weight of this "Kontext" has been trained for effects such as plot deduction with fixed subjects and advertising design.
However, limited by the number of samples in the training set, its current performance has not yet surpassed that of "NANO-Banana".
Nevertheless, its ability to mount other weights to assist reasoning is a potential advantage.
Moreover, when mounting the weight of "Another Angle", its effect in reasoning about the consistency
of multiple characters completely outperforms other similar models.
Thanks to the following developers:
WXL,Lyu Dexin,Eric, Longyiyu ,Opie
