Improved rendering through AI application

It is very interesting to be able to evaluate the hyper-realistic contributions of Artificial Intelligence (AI) to architectural rendering.

In this case we have exported renders directly generated with the Archicad engine. It goes without saying that the images were treated with such poverty, but this was the purpose of the exercise: to generate tentative or orientative images (almost volumetric) to be processed with some AI application called ARCHIVINCI: https://www.archivinci.com/

In general, it takes a lot of time to be able to record and detail each of the materials with textures that can resemble reality. The artificial intelligence interprets the materials and generates very realistic textures in terms of roughness, reflection and above all colour.

Archivinci works very simply. You have to select a type of tool that somehow guides you to the type of rendering you want to do. In our case, of all the options that the application has, we selected ‘Exact Render’ which aims to try to reproduce as faithfully as possible what is shown in the render while maintaining the integrity of the design, ensuring the accuracy of colour and texture in each render. 

Once inside you have to set a guide image, the accuracy/freedom of the texture balance that we give to the AI with respect to the supplied image, select the type of render (interior/exterior) and optionally we can play with a positive prompt and a negative prompt. These last two values are very variable and affect significantly the final result, so you have to know how to ‘feed’ correctly this kind of AI inputs. In our case it was almost nil, as when a lot of information was given, the results were generally somewhat misleading.

With this small experience we have much to comment on:

• The overall improvement of the rendering is significant, especially if we evaluate the time-quality variable.
• Although the quality of the final product is quite convincing, there are errors in some textures and some shadows do not match the volumes (especially in image 1 and 2).
• Something that caught our attention was the interpretation of certain ‘situations’ that the AI improved significantly: in one of the images, when interpreting that there was a swimming pool, it improved the water and put lining inside the pool, and in another it changed the furniture of a gallery for something much warmer and more familiar.
• Grass and trees, as well as the reflection of glazed surfaces are one of its main strengths, and it improves everything very quickly.
• When the AI does not interpret depths or volumes well, it invents a convincing situation. For example, see in image 1 the receded plane behind the gallery, which has little to do with the rendering obtained. Even so, the AI results in an incredible realism (which is not so much in accordance with the project presented as a guide).
  
• When we wanted to incorporate people or animals into the render, it has many problems to get them into the render. For example see the case of the dogs in the window of image 3 (the car and the baby with the balloon we put in Photoshop) and the people sitting under the gallery in the processed image 2 (if you zoom in the faces are very illegible).
• The application used has a paid option with a much wider range of tools and number of renders. We used the free version which is limited in quality and quantity, with a total of 3 test images.
• Incredible is the time it takes to make all these adjustments to the original image: no more than 10 seconds.

Improving hyper-realism in rendering using AI post-processing tools is very useful with incredible results. There are already rendering engines (such as D5 for a fee) that incorporate AI filters to improve the result of the images with excellent results. Undoubtedly, AI has arrived to improve and speed up our way of representing architecture.