A recreation of a complex of department stores in Poznan designed by architect Jerzy Liśniewicz. Built in 1968.
The model was created in Blender 4.0 based on available historical photos and drawings. Post-production in Adobe Lighroom. I used Photoshop and Stable Diffusion to inpainting people to give them a more historical look.
The car models are from the library of the Museum of Engineering and Technology, Krakow
This is quite a nice project. I’m from Poland so I am familiar with this architecture. Nice approach, quite complex and realistic. I’m curious though what is stable diffusion? Is it AI based application?
Thank you!
Yeah Stable Diffusion it’s AI based application. I used its inpaiting function to generate faces and clothes more in line with Polish reality
Raw render:
Stable Diffusion postproduction:
I did the same thing before AI in Photoshop but it took me longer and the result was not always that good. I wasn’t a fan of AI at first, but now it’s another tool in my work.
Thanks for reply. Yeah, If AI is used wisely and to improve something you can do yourself, then I suppose, it’s cool. I’ve never used it but I see that world is divided for those who use it a lot, and those who hate it. In my opinion Ai, if used- should be a tool, not an artist itself;-) But it a whole another topic 
Nice work anyways! I like Polish architecture, especially that from 50’ to 80’. We had really nice approach to designing at that time. It’s really nice that you delved into the history of this building to recreate it in 3d.
I featured you on BlenderNation, have a great weekend!
Thank you Bart 
What a lovely render. The fact humans look so natural in this environment it takes this scene to the next level.
I am about to work on the series of photographs that will need inpainting to be dome or 3D models placed in. I am torn between both prospects. On one hand, I have tried SD to inpaint but to my disappointment instead of a couple it inpainted random number of people, faces are disfigured and so on. 3D models, on the other hand provides less stylistic freedom, but than again it could be toggled in AI. Do you mind telling how you prompted/inpainted people and what SD engine did you use? Many thanks.
Thank you neb, I’m glad you like my artwork.
Regarding Stable Diffusion, I am just starting to experiment with inapinting, I have learned from this tutorial:
In my experience, the engine and prompts does not matter so much, the most important thing is to divide the inpainting into several stages and keep trying until you get a satisfactory result.
For example, when I was generating people, at first I generated the whole pose, and then separately the face.
Hey Mateusz,
Thanks for the tips. In your experience how good is the generated pose you start with?
For instance, Photoshop generative tool so quiet primitive. It introduces a lot of distortions all over the body, fixing is the same painstalking error-trial.
Regarding midjourney, I found reasonably good results, on this video. Multiply humans, poses, etc. Take a look :https://youtu.be/bRYFObeyzR0?si=xcyEIQs_7D-0yEJn
Have you actually tried inpainting/prompting ( i hope i use correct words) multiply people on the base image like; family, crowd , group of people, couple, couple and a child etc. so far?
Thanks
Sorry for the late reply, I have had a lot of work
In my experience if you’re generating from scratch it’s pretty bad. But if you have a placeholder, even a simplified one, in a similar size and pose and use simple, precise prompts then Stable DIffusion does quite well.
Yeah, looks really good
I tried, but the results were not satisfactory. In my experience AI performs best when it doesn’t have to generate too many things at the same time, because then deformations appear. That’s why even when I generate a large group of people in AI as in the case of this artwork, I generate people one at a time, this gives me more control, less deformations and better results
