This paper presents the proposed image encryption algorithm, a combination of variable-size block processing, a 6D chaotic system, and an automatically generated dynamic key. Bringing a new concept to design the new key based on extracted features by a Vision Transformer for the image, extracts the feature for unique keys that initialize the 6D chaotic system based on these keys. The new key hence is more immune to pattern replicated attacks after dividing the image into variable-size one-dimensional blocks determine by the chaotic values and arranging them into a 4*4 matrix, with each element of the matrix is a 24-bit value of RGB images. Due to the limited randomness of blocks, a second level of randomness is injected into the block by means of merging the rows in 1–4 mains including randomizing them, subsequent merging and leaving some rows unchanged. Then shuffling and rotating the rows are followed by row merging using a dynamic size of a row of 96, 192, 288, and 384 bits and controlled by the chaotic keys. The results of the experimentation prove the concept. The PSNR is 7.573 and 6.694, and the original image entropy is 7.7659 to 7.4617 and the entropy of the cipher image approaches the ideal one, which is 7.9991 to 7.9996. The value of the correlation coefficient is almost zero, meaning that there is no similarity between the original and cipher images. The NPCR is 100%, and the range of UACI is 33.88% to 33.98%, and the time taken is 1.2173 for the 512 × 512 image. This indicates the proposed encryption algorithm is highly efficient and secure, and it has better resistance against statistical and differential attacks, which makes it suitable for image transmission.