AGI - InstructDiffusion by Microsoft: better than SDXL?

Key Takeaways

hello community now I got a lot of notification today instruction diffusing by Microsoft research so here we have the archive pre-print September 7 2023 and as you can see here this is a very good depiction what this new instruction diffusion model by Microsoft can do you have an input image you can transform this image to a van Gogh style picture or you can in paint here some crown on the head of the elephant or you paint the pixels of the cheater here in blue or you mark the pixels of The Cat in the mirror or you indicate specific parts or positions in a picture or you do simply some detection of the objects in an image so sounds really interesting so let's have a look yeah Microsoft States here that they can handle all a variety of vision tasks understanding tasks segmentation task key Point detection generative task editing an image and enhancement in an image and also just to mention it outperforms prior methods on normal data sets and Microsoft state is a significant step toward some generalist modeling interface for vision talks advancing artificial intelligence in the field of computer vision says it is based here on the denoising diffusion probabilistic model you remember my last video where we went through all of the different diffusion models and they give you here a very impressive demonstration so they say hey look this is an image of a fridge with a lot of stickers and magnets and here's our model and we just say hey remove all magnets and all stickers and this is the result that we get and look at the other models so their model is the best amazing now I just said hey Bart by Google explain the denoising diffusion probabilistic model to me in technical details and then compared please to the stable diffusion XL model and here you have a short summary about the loss function the sampling procedure the number of layers and the image quality and the stability great but you know what really caught my attention here in this research paper by Microsoft is here this very subtle notice that the instruction diffusion model by Microsoft exhibits the ability of AGI to a certain extent so we have here some average General intelligent which is okay and then we have again here demonstrate the model exhibit a degree of AGI by leveraging the wealth of tasks yeah so instruction diffusion again with AGI and then you're not gonna believe it here in the conclusion here we have here that instruction diffusion and average General intelligent these are two labels they want really that you get this on a subconscious level and I think it's absolutely interesting how prominent here this AGI term is and of course they say that this model sets the stage for future advancement in the episode of artificial general intelligence in computer vision so you might say or you might think given this research paper by Microsoft this is an amazing new diffusion model by Microsoft and you know I was excited and I thought hey yeah look I finally I will do some very nice YouTube thumbnail where I say hey Microsoft AGI isn't this beautiful and then I don't know why but I had the idea listen let's try this model out to have here a performance on a platform Microsoft provide us here with a demonstration with an example to try here instruction diffusion out so here we go so here we are now here in the live demo when I say indicates the eyes with a yellow circle or go and as you can see we have here an official instruct diffusion by Microsoft so the result is not really convincing I see no okay so change the color of her eyes to Brown and go and we calculated an instruction diffusion gives us this result okay not really what I expected okay never mind this is just the beginning so let's go indicate the human skin we have a lot of cyborg mechatronic so what is the human skin make a segmentation of human skin okay um not really so change the female character to a male one and now we increase the guidance so I deviate from the default of Microsoft and now let's say a male character wow wow this is this is a beauty let me say this is a beauty so let's take another picture it's not working with this one so I say detection let's go with detection and you might say why because look here in the right hand corner detection this is exactly what I want and the result is nothing okay change the color from Blue to Orange now this is an easy task just take one color and substitute the color with a different color this should be perfect and here we have the result okay not what I expected okay now I say listen identify all object in the image a full segmentation so this is my command I deviate now from the original publication by Microsoft wow okay full segmentation is interesting so as I draw a yellow dot on her right eye so a single yellow dot on the right eye of the person in the image and we have the complete compute power of yes yes yes yes yes we have one we have one successful draw a Green Dot on their left eye after a minute of calculation we get now the result error it was too complicated this was a too complicated task okay Mark the pixels of the female character to Blue and leave the rest unchanged after almost a minute of computing wow this is impressive Microsoft impressive so let's change this marked a pixel here of the female character to blue because now we have a pure cyborg we have a human part and a cyber part now the okay everything it's get a little bit yeah look at this this is nice image editing add glasses so I have a simple human person just add glasses this is a simple simple task add the glasses let's zoom in a little bit and here we go yes oh wow okay I think after three minutes you get the idea so maybe we should say Microsoft is definitely on the way to some average general intelligence maybe it has not arrived yet