Hi Rizwan, great question. The text-based design guidance approach used in the paper is very much dependent on what can be visually identified from a bike image. Since the weight and fuel-efficiency are difficult to determine based on just an image, I think these objectives would need to be explicitly calculated. But with some advancements in multimodal learning, it could be possible!

Hi SunWoong, glad you enjoyed the video and the rest of the research at the lab. Happy to answer any questions you may have!

Great question. In terms of training time, I've found that traditional optimization methods and ML-based generative models can often be roughly comparable (though it depends on many things such as dimensionality and training parameters, or course). However, one strength of generative models is that they can be trained a single time in a conditional setting to then support numerous different target queries. In contrast, our traditional optimization methods will need to be retrained whenever constraints and objectives change. We are hoping to demonstrate such conditional training in an upcoming paper!

@@lyleregenwetter1588 Thanks for the reply. I have one more question. Traditional optimization techniques are used in various engineering design and optimization problems based on the nature of the problem, objectives and constraints. For example, scheduling problems in industries involve million of variables and could be solved by a highly scalable Linear programming problem or NLP, MILP etc. Also, lot of design problems that we tackle in industry lack prior data especially there is no image based data. Usually we do abstraction to arrive at a mathematical model for a given problem. To cut a long story short, my question is how practical are generative models in engineering design problems ?

@@levelupwithrizwan9371 I think your comment about the necessity of data is really the crucial factor here. Optimization is an excellent technique which can be used when no data is present. However, in scenarios where we are able to gather datasets, this data can give us several advantages. For one, data driven models may be able to come up with better solutions than optimization approaches and may be able to handle more complex problems (where identifying a set of closed form constraints to guide optimization may be intractable, for instance). Secondly, ideally data driven models should be able to identify and harness existing design expertise that is encoded into previous designs in the dataset.

I think this data reliance is the main 'impracticality' of data driven models but as more engineering design datasets become available, I see data-driven models becoming more and more practical. Already, some of these generative models have been successfully applied to real-world product design

@@lyleregenwetter1588 Thanks for answering. I would love to follow your research, probably send you a connect request on linkedin.

Hi Kai! Glad to hear you found our work interesting. Please shoot me an email if you'd like to chat sometime!

@@lyleregenwetter1588 Sure！writing now！