Choosing an AI Renderer

[JB Sties]

Hello,

I'm curious which third-party AI rendering software other users likes to work with. My goal is to create semi-photo realistic exterior images of my high-end residential models.

Thank you,

JB

Apple Studio / M3 / 96 GB / Apple Studio Display / 2 TB

Operating system used: Mac Apple Silicon Tahoe 26.3.1

[CosminF]

To be honest, I had some great results with ChatGPT. Better than anything I could do myself and for me it yielded better results than Gemini, SD or other AI's aimed at architects.

Some examples of GPT (plus) and he also did some great interior renders on more corporate environments which I can't show unfortunately:

Main points I learned:

- give it examples (links, pictures, references, product names) for each and every material
- give it  a clear starting screenshot, with as much detail in Archicad as possible because you don't want it to design on your behalf. 
- painstakingly prompt everything: scenery, material on back wall, what the roads looks like, how the silhouette of the hill is, time of day, where the sun is, what each material should feel like, weathering amount etc. And for best results tell him about the composition. Top right building corner is x, y, z.

- you can edit pictures once he generates the first one (zone select) but the more iterations you do the less and less true to the original it is. Usually I save the first picture, open up a new chat, start from "scratch" 
- tell it how the picture should feel, lights etc.

- tell it to keep the image proportions identical to the screenshot (sometimes it tends to mess with the perspective)

- tell it to ensure print ready quality and keep the photos high def (or if you want distance blur, tell it that).

- reference some of your main inspiration websites (while also keeping in mind, non-ironically, that AI companies stole a lot of artists art to train their data and that's not at all at the very least ethical - this is an important discussion to be had, for sure).

- use reference images (either to give it to ChatGPT but that can make him borrow unwanted elements) or use them to better understand the ambiance you are going for (like you would with a mood-board if you were to do the render yourself)

- post process it (affinity, photoshop, photopea) etc. Light balance, color grading, highlights, maybe add something to it "the old fashioned way". This way also the image doesn't automatically registered as "fully ai generated" in software trained to track specific AI generated slop.

- be wary of any artefacts and hallucinations that are not contractually covered. If you get random design choices there that you didn't want, there's also a risk that a client might understand them more and request them because sometimes the render is more convincing than a technical floor plan.

- keep in mind that you are also contributing to its training by feeding it your designs.

[Jp1138]

 No, I usually use Twinmotion. The lack of control of AI solutions make me nervous, but they are getting better everyday. Sometimes I use them to polish the twinmotion result, as you could you an image editor before, adding background, improving lighting, etc.