When Bloodborne Met Stable Diffusion
Ingredient 1: I've been playing Bloodborne again for the past few weeks. Of all the From Software games (although I haven't played Demon’s Souls and Elden Ring yet) it's probably my favourite. I think it has the most interesting story, a great combat system and it just oozes style and design.
Ingredient 2: I learnt about DreamStudio and Lexica also a few weeks ago. DreamStudio provides access to an AI system that generates images from text. Lexica is a search engine for those images with their corresponding prompts.
Ingredient 3: I have Soul Arts on my shelf. A book that was kickstarted and is full of beautiful art, (re)imagining different aspects of various From Software games.
Now let's see what will happen when those three things are combined, with some visual styling added into the mix!
A gothic lakeside town, sunset, reflection, bloodborne by from software, studio ghibli, horror, highly detailed, volumetric lighting, octane render, 4k
The first experiment above turned out way better than expected. Having browsed Lexica before, I had some idea about what instructions to give in addition to the actual subject matter that I wanted to see.
Gothic haunted village with fanged beast stalking streets in moonlight, volumetric lighting, oil painting, exquisite detail, 4k, bloodborne by from software, studio ghibli, horror
I generated three images and this I felt captured best what I intended, though not quite there, so let’s adjust the prompt a little a bit…
Gothic city street where a fanged beast is stalking in moonlight, volumetric lighting, concept art, painterly, exquisite detail, 4k, bloodborne by from software, horror
I am still missing the kind of fanged beast I’d like to see. In addition, two out of three of the generated images contained rather distracting visual artifacts, such as a street lamp hanging on empty air. Oh well. Now let’s see if we can do something a bit more intimate…
A blacksmith forging a weapon in an outdoor smithy on a hill, autumn, smoke, fire, volumetric lighting, highly detailed, sinister atmosphere, muted colors, in style of Leonardo Da Vinci
Here you can see many of the visual artifacts and oddities that seem quite characteristic to these AI-generated images, and were present also in some of the other images I generated: the person in the middle has face and limbs missing, the buildings in the background are just weird and wrong, and something horrible is going on with the right arm of the character on the right.
From what I’ve seen browsing Lexica, landscapes and cityscapes seem to generate best results.
Deep within the sepulcher traipses a caustic creature filled with bitter anguish, concept art, horror, bloodborne by from software, octane render, 4k, highly detailed, volumetric lighting
The first part of the prompt in the above image is almost word-by-word from one of the image descriptions in Soul Arts. The result is nothing like the image in the book, but I find it absolutely mind-blowing that an AI can create such results from a few esoteric words of text. It’s not only that the images would depict what is written, but it is the overall style in them that I find truly astonishing. All the examples above, and even the images that “went wrong” by having too many mistakes in them, are stylistically very consistent. For example, there’s no mixing of anime style cartoon characters and painterly landscapes.
This makes me wonder what skilled artists could do, when a tool like this will allow them to quickly explore different concepts, pick up the stuff they find interesting and inspiring, and then proceed to create their own work.
If you want your mind to be truly blown, go and see what has been created by using ‘Studio Ghibli’ or ’Dark Tower’ as one of the prompts. Some of the results are absolutely gorgeus and, if created by human, could only be called imaginative.
Have a lovely weekend! ;-)