A comparison image with 'Midjourney' which clearly shows that copyrighted content such as 'Mario' and 'Pikachu' is avoided in 'Adobe Firefly' of generative AI



Since the first model to be installed in the generative AI ' Adobe Firefly ' announced by Adobe on March 21, 2023 learns images from ' Adobe Stock ' and public domain content whose copyright has expired, It is characterized by not having to worry about copyright-related issues like existing generative AI. AI researcher Jim Huang posted a comparison of the image output by Adobe Firefly with the image output by the image generation AI ' Midjourney '.

Below is the image output by the prompt ' Deadpool posing on a car'. Deadpool, a copyrighted character, is output as something similar in appearance in Adobe Firefly (left). Deadpool is correctly output in Midjourney (right).




Next is the prompt 'Super Mario reflected in a puddle in a dim city'. Midjourney correctly outputs the requested character, but Adobe Firefly shows a dwarf who looks a lot like Mario. Mr. Fan commented, 'This interpretation is ... exotic (?)'.




This is how I tried to output Pikachu this time with the same prompt as above. Again, Midjourney accurately portrayed it, but Adobe Firefly gave it a slightly different interpretation. People who want to use generative AI commercially It is definitely better to use Adobe Firefly.




This time, the prompt: 'Photo of a large crowd of commuters in Tokyo. Eye-catching is a sharp-faced woman in red. Warm light, elegance.' Adobe Firefly has rendered a red-clad person who stands out from the crowd, but not the facial features. Midjourney, on the other hand, seems to depict the prompt more or less accurately. 'These prompts are heavily optimized for Midjourney, so they can be unfair,' Huang added.




This is the prompt 'urban architecture of abstract fractal, circular, mosaic'. Although the direction is different, both seem to correctly reflect the prompt.




This is what was output with the prompt 'iOS application icon, skeuomorphism style SF planet landscape'. It was entered to confirm that ``Does Adobe Firefly understand the instructions for the application icon?'', but it seems that it is not good at this.




In images output by AI, the shape of 'human fingers' is often strange. Adobe Firefly also has this problem, rendering the human left hand fingers slightly blurry in the output.




Finally, the result of trying the prompt 'Red Ferrari F40 surrounded by dandelions at Lake Seealp ' is like this. Seealpsee is a Swiss lake surrounded by mountains, and Midjourney seems to reproduce the mountains in the image. Adobe Firefly is a little more plain and has a slightly more illustrative image compared to the photorealistic Midjourney.




Mr. Fan said, ``Because the learning sources of Adobe Firefly are limited, data curation is very conservative and performance may be degraded.'' ``Companies that must secure the copyright of the output image. For businesses, Adobe Firefly could be a big plus, and for companies willing to sacrifice quality for legality, Midjourney is a less attractive option.'

Related Posts:

in Software,   Art, Posted by log1p_kr