Take a look at this, this has the possibility to be a game changer in the making.
Just imagine what has happened in the last 8 months, now we have the control net, next will be consistency IMO.
3 Likes
Hello Holo,
I think that the idea of using AI and or ML ( the distinction between the two is not particularly clear to me) for rendering seems a good fit. Thank you for bringing this up. Informative.
Andy
2 Likes
Beyond any doubt yes.
In the future, I’d like those deep learning tools to be more assistive and less your competitor. What the model spits out in the video does really only very marginally correspond to the input, which makes the input in these examples basically irrelevant, since you could have done something similar with a less involved text prompt.
I mean this is a good tool for people that are deprived of any creativity and will to manifest their own vision into a project, but for everybody else it should be seen as a threat.
It’s so boring and generic…
Quoting human lyricist Flavor Flav: “Don’t… d-d-d-don’t believe the hype”.
2 Likes
when something seems to become irrelevant or threatened to be it is a good time to abstract and to rethink what really necessary is and questioning life, purpose and joy or ways to enjoy including what competition is and how to use competition to either strengthen and equally ease our society becomes imminent.
i understand all the scarecrowing going on against the ai, i have been participating either… but well as written above maybe its time.
it often reminds me of a very old Rhino Plugin that made designing jewelry pretty simple. it worked fine as long as you didn’t mind it making decisions for you and expressed in very aggressive manner: all jewelry looked about the same in the end.
Now being able to create jewelry that EXACTLY what my clients want, is the reason why I still have a job and clients. It’s also the reason why I still work with Rhino and not more dedicated programs that make half the work for you in no time, but always the same way. Otherwise you could just download some "ready to print 
" files from sketchfab,cgtrader or other 5$/design website.
These facts won’t stop me from using AI or those websites in cases it makes economical sens. I’m working for an industry that is still intimidated by 3d printers and SLM technology and still makes circlepacking by hand, putting circle after circle on a surface… I truly believe AI will assist us, not replace us. At least not entirely in conceptual work when the demands are very specific… Repetitive work though…
3 Likes
I imagine having placeholders in the 3D model that works as inpaint masks, and they are tagged with keywords for the AI to use for content creation. To me that would be a great way to add foliage, trees, people etc.
I have been testing out inpainting in Stable Diffusion but haven’t gotten any good results. Have any of you?
1 Like
Well - that was the story with Keyshot/Hypershot.
When it went out, all the rendering at our design school immediately started to look “Keyshotish” - and it is quite difficult to “Keyshot-out” the rendering in Photoshop or so, once it it finished.
But I would say that with AI renderers it will be different story - yes, it is correct, that with prompt too simple, all the pictures from Midjourney looks kinda same (the same style).
But with prompt that is complex enough (AI is strongly “forced” to some specific style), it is very difficult to tell if it is AI render or very good Photoshop, or even some Illustrator artwork.
Especially StableDiffusion is very good in mimic very different visual styles.
The disadvantage of AI renderers today is that they make very nice picture - but only if you need one (picture of the object).
If you need more pictures of the same object (different angles etc.), you never get “the same thing”.
But maybe in future that will be solved.
Anyway, YES, I think AI is future of rendering.
Maybe it will not make “on click renders” in the end, but the help of AI can be massive.
I remember days of old V-ray in Rhino 5 - that was terrible thing with dozens of obscure hypertechnical parameters to set, that were totally irrelevant (like “photon size” or “initial photon grid” - who cares what is the photos size suppose to be? I want nice picture in 15 minutes! Hey SW, set something reasonable there!)
It is good that renderers made big step in accessibility (mainly thanks to physically correct unbiased rendering that removed those obscure parameters) and AI can make another step in accessibility that even more people can push out nice pics out of the 3D package without need of study of rendering for weeks or months.
2 Likes
Thanks for evolving your interesting thoughts and experiences.
I think it’s very true what you say here, just I think when people are refering to AI it is to refer to something that makes all for you with very little number of clicks. (like Keyshot that flooded the place with it’s flat renders because it is so simple to use; like Matrix that made a ring with 3 clicks.) These comfort solutions often have a huge success so will AI rendering for sure.
Ideally AI would set all values in a classic renderer like Maxwell/Bella and we’d still have the possibility to adjust what you consider not well chosen for your specific render. Maybe some kind of “dynamic default values”. This would be the best of both worlds, would give entire liberty combined with time-saving help from AI and in the end it would become much easier to just change one value in a working render and see what it did. this would be a great source for learning these settings, too. I’d embrace such a solution.
Key differences between Artificial Intelligence (AI) and Machine learning (ML):
| Artificial Intelligence | Machine learning |
|---|---|
| Artificial intelligence is a technology which enables a machine to simulate human behavior. | Machine learning is a subset of AI which allows a machine to automatically learn from past data without programming explicitly. |
| The goal of AI is to make a smart computer system like humans to solve complex problems. | The goal of ML is to allow machines to learn from data so that they can give accurate output. |
| In AI, we make intelligent systems to perform any task like a human. | In ML, we teach machines with data to perform a particular task and give an accurate result. |
| Machine learning and deep learning are the two main subsets of AI. | Deep learning is a main subset of machine learning. |
| AI has a very wide range of scope. | Machine learning has a limited scope. |
| AI is working to create an intelligent system which can perform various complex tasks. | Machine learning is working to create machines that can perform only those specific tasks for which they are trained. |
| AI system is concerned about maximizing the chances of success. | Machine learning is mainly concerned about accuracy and patterns. |
| The main applications of AI are Siri, customer support using catboats, Expert System, Online game playing, intelligent humanoid robot, etc. | The main applications of machine learning are Online recommender system, Google search algorithms, Facebook auto friend tagging suggestions, etc. |
| On the basis of capabilities, AI can be divided into three types, which are, Weak AI, General AI, and Strong AI. | Machine learning can also be divided into mainly three types that are Supervised learning, Unsupervised learning, and Reinforcement learning. |
| It includes learning, reasoning, and self-correction. | It includes learning and self-correction when introduced with new data. |
| AI completely deals with Structured, semi-structured, and unstructured data. | Machine learning deals with Structured and semi-structured data. |
1 Like
How I see AI rendering for the future is image->text->image with inpainting masks from 3D objects.
I would love to add people, furniture, vegetation etc as generic objects with text tags that are interpreted in the inpainting pass.
Hello Ben,
Wow - Thank you so much* for your reply of : Clarity; Speaking of, as your being a jewellery person is AI / subsets already being used in video / rendering of jewellery ? [or simply a picture]
Is the visualization of jewellery that what we sometimes see on screen or in print a render / model of a jewellery design or an actual tangible good ?
For the visualization and for the determining of the quality of gemstones it would seem like AI / subsets would be a good fit. Particularly when the jewelry / gem / gemstones are visualized moving or able to be moved in a 3D space.
Below are some examples of visualizations that I think are nice at Graff’s Instagram site in their reels section:
Or here [ in the form of a render / 3D model or tangible good ? again, I not sure]:
So the wonderment of it; if I am looking at a visualization of a tangible good or not. And irrespective of this, the utilization / potential / further utilization of AI for the visualization of jeweler is quite interesting.
Thank you,
Andy
- to and everyone on this thread, started by a great question.
More tensor cores, less RTX cores…Soon.
Well, I cannot speak for everyone, but as usual, there are several quality levels of renderings:
-
For some Jewelry vendor web site, surely AI will be used or is already, but from my point of view the results are not too far from using Keyshot.
-
The problem with Keyshot and also AI will be that they simplify things to get quicker results. whatever you do, the maxwell equations must be calculated if you want all physical effects respected in gemstones (IOR, transparency, luminescence, pleochroism, idiochromatism and so many other effects) and therefor no quick solution will be possible if you want it physically correct. e.g. I’d made this render 2007 where I applied physical thickness of around 450-750nm, not even speaking of the excessive rendering time:
which is still better than keyshot in 2020 in regards of physical correctness, but it’s a question of need. Noone will pay money for the hours I spent on this.
Let’s say I got a ring that is produced just once and is in a price range of about 3000$ including material , production and margin. It will be difficult to bill 300-500.- for a perfect rendering. In this case you go with an AI or keyshot or simply rhino render or cycles solution. On the other hand side there is no need to have the gems reflections perfect.
If you have a piece that will be presented to a final client with a budget of several hundreds of thousands, it’s a complete different story and perfection is all that counts. No way this client will see some quickshot rendering.
So to conclude a bit:
- Graff: 100% rendered with low effort. mainly visible because of the surfaces finish.
- Verlas: 100% rendering, mainly visible in the way too proper graving. Also the metals reflections seem extremely flat/unpolished, which indicates a workaround to avoid only white surfaces (as polished metal is a mirror essentially and with white background you’d see nothing. Clear indicator of a rendering done by someone not knowing much about photography settings or a program that doesn’t allow such settings like AI or keyshot)
- even from real photos sometimes you cannot be sure wheter it’s been rendered for different reasons:
- images are often compressed to jpg format, not respected the pixel shift ratio when scaling down or up the image which leads to tiny deplacements mainly make gem reflections blury or other minor technical errors.
- everything is being photoshopped after the photo is taken. There are examples where I’m sure it’s a photo from the physical piece, but you’d think it’s a high end rendering.
However, both examples don’t need more details, as the final clients won’t bother if the reflection of the stone is a simpler version of reality. The idea and general appearance is represented well enough to be understood. - the most common error is the missing depth of field. usually to take a photo from a small piece like a ring, you need a macro objective on your camera, which physically has a tiny depth of field. e.g. you would not be able to see the engraving sharply as the gems in front if 1. it was no rendering or 2. there have not been taken two photos at least with different fields of depth and photoshopped together.
In all cases you should never (NEVER NEVER) estimate a gem’s or jewelry’s quality by a photo. It’s way too easy to manipulate digital data nowadays and with jewelry we usually are quickly talking about sums that attract scammers.
So yes, AI (or any equivalent quick render solution) is extremely interesting for us jewelers if you work with final clients, but again, letting make something everything means you don’t know what’s been done.
The only thing bothering me in this video above is that main shapes are just changed. If I’m to spend the same time controlling all details of the AI rendering to be correct, I’d rather make it from scratch myself:
here, left and right are very far from being the same thing and if I had drawn the left, I don’t know how I’d justify the difference between preview and the physical object in the end.
Although I was not very clear, I hope I have still given you some insights and hopefully answered your question. If not, don’t hesitate to relaunch the question.
Ben
1 Like
Keep in mind that text->image was launched under a year ago, and image->text->image is brand new, so if this continues to develop in the same speed then all your worries should be dealt with within a year or two IMO. For AI to be able to assist in rendering we need consistency and an advanced and evolved control-net that handles any shapes.
I did a quick 5 sec test on an old rendering of mine and if this evolves it has potential.
I’m not a skilled prompter, so if anybody is then please run some tests and share your results, it would be cool to see what can be achieved already now.
1 Like
you’re absolutely right, I am sure it will extremely evolve now it’s become very popular lately. I was talking about this very video and the “not-yet-there” character of this very solution for precise preview of the object itself.
however, if I dare to say: in your example there’s a little missed interpretation of the canny-generated edges. (also canny has, like all edge detection algorithms, its weaknesses and your image is relatively convienient for canny)
the geometry of the car itself has changed. what was initially a total reflection of a light source has become an edge and has been extended from the initial canny line to the bumper, which is very wrong if you showed it to a client and fabricate the model from the left… Now knowing that we can hardly understand what the AI is doing, it might be slightly difficult to correct that occured problem without any knowledge on the process.
oh and I see just now, the lights of the car are also completely wrong and make no sens except on a planet where we have 3 suns.
I very much prefer your old sexy car rendering BTW 
1 Like
I’m working in automotive, that is probably quite different from Jewelery, and from my point of view:
“I don’t know how I’d justify the difference between preview and the physical object in the end.”
You will simply not.
When we are designing and modeling car, we do thousands of renders during the 3-5 year development (some of the are just screenshots, some of them more elaborated for internal/decision purposes etc.)
And for every purpose (as not each one is presentation for customer) you need different type of rendering - even now we us e at least 3 different SW/renderers depending on what we need, how fast and for what purpose.
Right now, AI can’t be use for final product renderings that customers will see.
But that is not a problem - we do thousand times more renderings during the development, that customers will never see (thus we don’t need to justify it to anyone).
But even now, if the tech stopped today, there is a big field for AI renderers, for let say “research” or “brainstorming” phase of project.
Where you need only the main volumes to get right (there the text>img is not good, as it is very difficult to get the desired sizes and proportions of objects out of it).
Martin
3 Likes
not really, your field is just much more complicated and regulated and therefor take more time.
thanks for the insight! your text was way better towards the topic:)
I think you are absolutly right with what you said and all like you, if I work with professionals, it’s very different. There I’d rather use whatapp videoconference filming my rhinoscreen with default material:)
It really depends on what we need and indeed for final products it’s not fit. for many many other applications AI is clearly a game changer and will much more be once 3d models are fed to ML’s and not only text or images.
Thanks @Holo for bringing the topic up. this is a extremely interesting thread!
1 Like
I did some image searching on the topic and here are a few finds to inspire the imaginarion of what might come:
1 Like
