Suppose that I put this prompt into Nano banana to create an image for a scientific paper
create
Figure Concept 1: The "Privileged Information" Diagram (Cross-Modal MKD)
Goal: Address the reviewer's specific suggestion about "adding/removing one type of image data (like T2, T1w)."
The Story: Show how a Teacher sees everything (T1, T2, FLAIR, T1ce) and teaches a Student who only sees one thing (e.g., T1) to see the "invisible" tumor boundaries.
Visual Layout Plan:
Left Panel (The Teachers): Display a vertical column of actual brain slices from the BraTS dataset for a single patient:
Image A: T1-weighted (Anatomy)
Image B: T2-weighted (Edema/Structure)
Image C: FLAIR (Edema is bright)
Image D: T1ce (Tumor core enhances)
Right Panel (The Student): Display the Student network receiving only the T1 image.
The "Magic" (Distillation): Draw connection lines showing the Teachers transferring "Knowledge" (not just labels) to the Student.
The Result (The Neuroimage): Show three segmentation comparisons:
Baseline Student (No Distillation): Fails to segment the whole tumor because it can't see the edema in T1.
MKD Student: Successfully segments the tumor core and edema (mimicking the Teacher) despite only seeing T1.
Ground Truth: The expert annotation.Gemini output this image
which looks very undesirable with too many redundant extra details. Writing prompt to edit this image would take a huge amount of time, and after a few rounds, Gemini will just get confused and everything become a mess. Therefore, I would suggest to use your own hand-drawing to help Gemini understand what you really want.
So, in the same conversation in Gemini, I uploaded my own scratch and use prompt Create an figure based on this instead, i.e.,
With only one shot, Gemini give me a very desirable result
Discover more from Knowledge sparks
Subscribe to get the latest posts sent to your email.