Umění v AI – MidJourney X DALL-E 2 X StabilityAI

Cool a přehledné srovnání. I když myslím v závěrech dost zavádějící.

Každý model má svá specifika a je ideální jim prompt přizpůsobit. MidJouney exceluje u dlouhých „konvergentních“ promptů, ale není zde mezi testovanými jediný takový. Dalle-2 zase dokáže napodobit mnohem lépe umělecké směry, než je zde ukázáno a v portrétní malbě u mne jednoznačně vede. MidJourney je zase ideální na krajiny nebo architekturu s neskutečnými detaily a osvětlením. Také používám i jiné modely, ale sám nemám zkušenosti se StabilityAI.

Každopádně žasnu nad tím, jak se v této oblasti AI posouvá doslova každý týden!

 I have finally been able to test a new AI: StabilityAI. I was really looking forward to it after the creations and the comments I was seeing in social media and forums; and the truth is that it offers very interesting options. Let’s compare them!

 Attached you have a comparison of MidJourney, DALL-E 2 and StabilityAI (in this last one I have used the k_lms sampler, I will explain later). I have divided it into two areas, Convergent (when we ask the AI something with a lot of detail and specifications) and Divergent (when we give it very few guidelines and let it explore). These in turn are divided into two:

• Convergent // Real. When what we ask it to do is something that could belong to the real world.
• Convergent // Imaginary. When what we ask for does not exist as such, they are new creations.
• Divergent // Concrete. When, although we let it explore, we are asking for a specific result.
• Divergent // Abstract. When we are asking it to give us sensations and imaginary, nothing concrete.

I think the results speak for themselves but I am ready to draw some conclusions that I find interesting:

• For imaginary and sensations, MidJourney is unbeatable.
• For original, different, „creative“ results, I think MidJourney is the choice.
• For realistic results (in general), I think DALL-E 2 is the best option although depending on the case, StabilityAI poses interesting results. I would try both. MidJourney fails because it is not capable of making coherent creations with a lot of detail. It looks like it but it is not.
• For humans, if what we want is a result with photographic realism (as close as possible), DALL-E 2 and StabilityAI are again the options. If we want it to be done in a specific artistic style and not expecting it to be perfect, MidJourney is again the choice.
• I think StabilityAI is in the middle of DALL-E 2 and MidJourney, being more artistic than the first one and more realistic than the second one. I think it can be a really powerful option to consider. In addition, it introduces a new variable: you can choose the method (sampler) with which you want it to generate images. For the comparison I have always used the same but changing the sampler depending on the creation can provide really interesting details.

 We’ll have to follow them closely and keep playing around!