Nutlope/roomGPT: open-source clone of Interior.AI | Hacker News Reader

Nutlope/roomGPT: open-source clone of Interior.AI

github.com

151 pointsswyx32 comments

Does this have anything at all to do with GPT? Would be sad to see this kind of naming scheme become a trend...

It's a Generative Pretrained Transformer, same as all the others.

http://ww01.interior.ai/ is a spam domain parking page. Perhaps the submission title should be changed to InteriorAI[0]?

[0] https://interiorai.com/

It gentrifies even the occupants! https://imgur.com/a/CduMyPo

Was looking to make something like this, but glad its out there already. The real value though would be to make the app aware of furniture and lighting inventories and cost of paint and whatever else goes into a room. Then after generating in a style, maybe you can say "same thing but cheaper options".

We'll get there eventually...

I'd invest in that

The generated room in the example is fairly hideous.

The example has so many bizarre artifacts it makes me think this does not work very well at all

i feel that way about everything that has been showcased as state of the art AI

Then you haven't been following the recent developments of stable diffusion. It is mind blowing how good the quality gets if you are using stable diffusion + ControlNet.

I guess the reason this particular example does not work is that a descriptive prompt is missing and / or functionality like inpaint is not available in the process. You can easily fix most of (all?) the artifacts using inpaint.

Can it do hands & fingers now?

Stable Diffusion is amazing, but I've never seen it reliably generate realistic hands. The examples in the ControlNet repo show a chef in the kitchen with 6 fingers.

Oh well that one you can edit with photoshop, otherwise it do creates way perfect images.

Yes you can. Came out a few days ago. Open pose has already support for it. You can not only render realistic hand and feet but also customize the hand gestures. There is a blender plug-in which let's you position your character pose including hand gestures in 3D and then render the needed controlnet masks. It works amazingly well.

i've been watching and i just haven't drunk the kool-aid. there are things that are just off about the results, but i'm stable enough to not be diffused by the hype.

I guess you can always hate on things which aren't production ready yet. Considering the progress made in the last month's and the speed things keep improving it will most certainly have impact on how the industry works.

That being said, why are we only talking about stable diffusion. What about openai whisper? I have never used a better speech to text software and it is multilingual.

It can be enough to give a general idea of a style change though.

Watching the source code, I only see this as a Web UI for paid model service replicate.com

This is a good start, but it’s unfair to call this a clone. The generated images are filled with strange artifacts that you don’t see when using interiorai’s product.

That's just the final fine-tuning output.

I'd say it would be easy to add the layer required to improve it by yourself. It's open source after all.

Hassan has been doing great stuff so I'd at least give him the benefit of the doubt

Sorry for the stupid question, but what do you mean by this? Adding a LoRA?

what's a LoRA?

Low-rank adaptation - an efficient way to fine tune some types of model. https://huggingface.co/blog/lora

I meant playing with SD vanilla, you can add LoRA/TI/Dreambooth for custom stuff, but for generics, you need to improve your main pipeline in order to not show bad results.

You can also check aesthetic score of results, or hide results w artifacts etc.

My bet is that as a demo tihs has low steps (cheaper replicate tab), and with a different diffuser (some work better for photoralism than others) and more steps results would already improve greatly

Check out the code, and improve it thats what I meant.

Worse than artifacts, it has things that make zero sense in terms of functionality and placement. Unless we're talking about low-grade fiction (on a budget, where authors can't or won't spend time working out the fine details), interior design is about function, not just nice-looking pictures.

Lots of clones out there, this one in spanish https://decoria.app

And this one as well https://stageinhome.com/#pricing

This is just an api wrapper calling replicate.

Exactly

This is a rather poor product.

Why would you spend energy commenting this? Why not add something to the conversation or give some constructive feedback.

There are plentiful opportunities for improvement.

I wanted to express disdain for the couple of minutes I lost while testing it out. I stood up to take a photo of my living room and found the results to be abysmal.

Reply on news.ycombinator.com