Kitchen Renovation ideas animation using Stable Diffusion [video]
youtube.com
Using stable diffusion to generate and walk through different kitchen renovation ideas.
youtube.com
Now I know that this tool is just a “starting point” for using as ideas, but I have seen way more fabulous (and sane) kitchen designs en mass on houzz, dezeen, basically any online blog. The technology is cool, but it seems limited. As in we can only get weird amalgams of human concepts. All this to say, I work with architects professionally, and I do not see this as something that will cause their field concern in the near to mid term. These programs cannot make drawings with structural suggestions, floor plans for the layout they produce, or anything else that a person pays an architect for. Just my two cents
What exactly does that mean? I bet most people would fail to tell things apart on a blind test comparing kitchen designs generated by humans vs. ai (not counting the "transitory" images with blurred corners etc.). Same thing with concept art.
I'm not talking about "transitory" frames. I'm talking about things like the bottom of the top kitchen cabinets frequently being missing, top beams being at different levels in an impossible "MC Escher"-like configuration.
I definitely could see this being useful in plugging in different "styles" to decide what you like, and then taking that to a designer or architect to build out a real design, but anyone paying attention should easily be able to identify the generated images.
Somehow I doubt those blogs have the same floor plan as my house or any easy way to filter for things that would fit in my own kitchen. Don't underestimate the power of seeing a design in your exact kitchen/space. What looks good in one layout isn't guaranteed to look good in another.
> All this to say, I work with architects professionally, and I do not see this as something that will cause their field concern in the near to mid term. These programs cannot make drawings with structural suggestions, floor plans for the layout they produce, or anything else that a person pays an architect for.
Ok but that's not the point, at least it's not for me. I'm not trying to replace architects, I'm trying to figure out what I want so I can take it to someone who knows better and say "I want this". They can kick back anything that's not going to work structurally but I kind of doubt this is going to be a big problem as, not to be an ass but, I have a brain. I know where the cabinets can go (the same place my existing ones are) along with my sink/oven/dishwasher. What I'm looking for is designs/colors/etc and how that would look in my kitchen, I'm not designing a kitchen/house from an empty plot of land.
As with all Stable Diffusion (and most AI stuff, see CoPilot) the goal (or at least my goal) isn't to replace the designers/architects/programmers/etc but to automate away the parts that waste a bunch of time. For CoPilot that's boilerplate or similar code blocks, for this it's letting me see designs and (hopefully) tweak or re-gen parts I don't like until I get it close enough to take it to an expert. I fully understand the "kitchen remodel" tool doesn't exist today but it's not hard to imagine what it would look like with image input, in/out-painting regions, and a nice GUI wrapped around a Stable Diffusion core.
Someone with zero design experience could run this over a photo of their kitchen and realize that their exact real world kitchen would look nice with a different set of paneling, or a different set of chairs, or a larger window, etc. Obviously they have to consult with an architect or other expert to see how feasible larger renovations are, it's not meant to replace architectural work, but there is plenty of value already added as an inspiration tool. In fact it might lead to a lot more work for architects, if they can use a tool like this to show potential clients a bunch of inspiration for how their kitchen could look, without doing any actual architectural work up front.
Once that part was done, then the designer could use those bits and pieces to identify what you like best and cross reference that with what is available and what is possible within your budget and get you the result that you would be happiest with.
With that, the actual designer is still a very important piece of the puzzle and this does very little to replace them.
When the AI blows away support walls and rearranges things that can’t structurally work, its utility is limited too. It can also dream up sinks/fridges/ranges that don’t exist with sizes that aren’t realistic either.
I’m supportive of the general direction but in a model that’s built for it where all the real constraints can actually be accounted for reliably. That would allow you to go beyond looking at pics for inspiration into renderings you can flip through easily. Bonus points if it can approximate cost, including plumbing and electrical in rearranging things for where plumbing and electrical already are.
...and it's already led to actual change: https://twitter.com/MikeBradleyMKE/status/157083982421467545...
The conversion pipeline for things like this today relies on ads, word of mouth, and expensive sales processes. Imagine a homeowner uploading a photo of their kitchen and being shown hundreds of possible redesign styles. The homeowner can choose one to start a conversation with a designer. That pipeline will convert way more people than the status quo, and the industry as a whole will grow.
So much of the reaction and subsequent reasoning about such tools is bounded (predictably, inevitably, naturally) by human cognitive error—in particular, our total collective inability to reason with non-linear extrapolations.
We "know" this is a starting point, but even the linear line from "nothing like this existed at the start of 2022" to now, projects into a "what exists in 18 months" which is itself an effectively instant timescale (residential A&D projects here in SF are often 3-year endeavors...)...
...what this looks like for a loose interpretation of "this," is literally unimaginable to us.
One technical thing: I was just telling someone 30 minutes ago, in the lats week, my #aiart timeline and feed on Twitter started saturating this week with extrapolations into AR/VR applications.
The toolchains which will exists in this domain prior even to the launch of Apple's long rumored AR goggles are already eyebrow-raising. Assuming there is a path to deploy experiences onto Apple glass, there's going to yet another infusion of energy (cough money) into the precambrian explosion going on around ML/imagegen etc.
Don't make the mistake of thinking flaws in proof-of-concept first-passes like this (which is already IMO quite compelling for ideation), are going to be a permanent fixture or blocker.
If you review the evolution of imagegen over the last two years, you can get a taste of how ephemeral most such limitations are...
...and the egregious ones tend to attract a frenzy of work.
To put a point on it, I think you're OK saying "near" term, but "mid"... absolutely not. This is just one more place where disruption is clearly visible on the horizon.
In some way, stable diffusion works like the architect mind that has seen more kitchens (in this case) than the average person and somehow (consciously or unconsciously) proposes a solution based on that. Our edge - for now - is knowing what is feasible/economical.
You could do this for any sort of web / mobile / whatever UI, designing houses, cars, any object really, fashion, etc. It's all going to be so much faster.
That's exactly what the average creative professional does. As an anecdote, being intellectually honest, most of the architects I know (and I know many as an housing architect myself), would not do better than this tool does now - "creatively" speaking - let alone in the future.
But still, StableDiff has been amazing in my view. Prototyping anything this quickly is going to be a game changer. As someone who doesn't think in pictures well, this looks to let me mock things up so incredibly well. I am so hyped for the refinement of the tooling in this space.
I wonder if those could be implemented as vector addition in the latent space.
Take a picture of a room then use img2img in stable diffusion automatic1111.
Set denoising strength to a value between 0.3 and 0.6. Lower than 0.3 and not much will change, higher than 0.6 it tends to start changing the whole room.
It can be used without a prompt or use a prompt to attempt to describe the room, like a "a living room with a grey couch a white table walls with wallpaper" then change parts of the prompt to morph it in desired direction, like add a chair something or add the name of art styles or "by someartistname". Just having a prompt that somewhat describes the room and pressing generate many times will also just create a lot of random alternative interiors, like in the video. If you want to create a video I guess you'd use deforum, but I haven't really tried that.
The stuff documented under img2img alternative test here can also be useful if you manage to dial it in properly. This is for keeping the image more stable and getting more control of what you change. https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki...
I own a SaaS that offers online 3D configurators for outdoor structures [1] since 2014, and tried something similar as a UX experiment a couple of years back; here is the gist of it:
We offered 20 different variations on the screen at the same time, and people had to gradually fine-tune their preference by clicking on the one they liked the most.
I tested this in the field, and it proved to be very confusing to the end-user, unless you highly constrain the amount of parameters that can change at the same time, and describe to the user what changed and go step by step.
In the end I decided to ditch the whole idea in favor of a hard-coded step-by-step wizard where users can adjust the relevant parameters themselves one by one, which tends to work the best for the majority of the users.
=> My suggestion for you would be to do something similar: highly constrain the potential changes, and guide your user step by step. If you manage to do this, I think you might have gold in your hands.
I would also love to challenge the naysayers who say the generated images don't make sense: in my opinion it will only be a matter of time before someone starts training or hardcoding a classifier that invalidates "wrong" images. (Especially if one would be able to generate a reasonable 3D representation of the image without too much effort.)
As someone once said in a random AI video on the internet: "Dear fellow scholars, I wonder what we will be capable of just 2 more papers down the line"
[1] ** edit ** removed the URL, check my profile if you really want to know...
1. Simply remove the ceiling beams.
2. Extend the room and place another window into the wall.
3. Replace wall with window.
4. Cut room in half and replace one part with garden.
5. Open room to garden and slowly extend ceiling into garden.
Stable Diffusion opens a lot a avenues and yet it is somewhat "restricted" to what it "knows/understands". Put differently, if cars it exposed to is black can it paint them pink ? What about even higher level construct ? It can not innovate can it ? Not yet ?
Anyways, fascinating.
Very few people in creative industries are truly pushing that industry into new frontiers, and that is actually appropriate. Exploring outside the frontier is actually full of failure, and far less likely to produce something most people want.
I honestly think that people operating outside of the frontier of established creative trends technically have a completely different set of skills, and a completely different job altogether, than people who create within the boundaries.
However, we should not neglect small incremental innovation that is created by combining previously known knowledge in new ways.
And only show me lights that really exist and are for sale, etc.
There are a lot of changes in the windows and other stuff that is difficult to modify in real life. It would be nice to be able to restrict the changes, but it's probably difficult to "explain" that restriction to Stable Diffusion.
Same issue though, it may suggest you remove your ceiling or move a door.
It’s an apt analogy for chasing perfection and being trapped on the hedonism treadmill.