Image Generation With OpenAI API Theory | Complete OpenAI API GPT Python Tutorial Part 9

Key Takeaways

what is up guys welcome back welcome to this tutorial in this tutorial we're going to start with image generation um we're going to um do image generation using DOL 3 and do 2 as well and basically we'll be not just generating image we would also be editing image and we would also be seeing how we can do variations of you know the images that we let's say have so we will be doing all these things um and in this tutorial particularly we would just be covering the theory part in the next tutorial we would start with uh coding and and then later on we'll have a project on my UD course as well so let's start with this thing I've just uh made a small presentation for you guys so let's start with that okay let me make this thing bigger there you go perfect so as for the image generation there are three methods as I've said before the first one is basically creating images from text prompt so you give a text prompt and you know it creates an image and you can use D 3 and D 2 both of them for this thing however uh for the other two you can only use di 2 um and the other two are um as you can see on screen uh creating an edited version of the images uh by having the model replace some area of the pre-existing image so basically they have mentioned this thing using a uh photo in here you might have seen this like in uh some of their um you know their I believe um when they you know released this this thing and there was like a there were a lot of videos in which you know people tried to experiment this this thing so as you can see this is a regular image we masked um you know some area and we gave a prompt uh as you can see this is the prompt then it g it would be giving you you know like you know this thing on that prompt so that's the idea of editing the images you don't have to go ahead and you know do a lot of editing basically it would you know edit the image for you and then creating variations of the existing image basically it would take an image maybe it might take um you know if you want to give it it can take a prompt or you can you know leave it like that and you know you can generate variations of the existing image right obviously these two things as I've said before are only with d 2 right cool next As for the generations that the images that are generated with Del 3 the standard size is you know this thing um basically this is a square I believe this is a landscape and this is a portrait um so yeah you can generate all those uh three the default uh in Del 3 is standard quality but you can also always you know give this as a key uh quality as HD and in that case you know you will get a better quality you know images with d 3 you can only request one image at a time um so basically you cannot have um you know multiple choices of images that uh you can you know generate uh with Del 2 that's not the case you can always you know change the end parameter um and that you can you know do with the code and with the API and basically you can generate up to 10 images at a time so you can have one prompt generating 10 images right uh Del two Del 3 automatically rewrites The Prompt for safety reasons so basically uh this is something that they say on their on their official page and they say that Del 3 would do this as you can see in here uh now takes the default prompt provided automatically and rewrite it for safety reasons um you know for so this is you know something that they do and you can always check that uh with this parameter so basically you can once the you know output has come you can check it in inside this in which you know you would be able to see what actual prompt was you know given uh and they're saying that this is currently it is currently not possible to disable this thing um right so that's the issue cool going ahead there were there are few python specific things now before we go that um before we discuss python specific things let's just discuss you know uh some of uh these things here so as you can see this is an example here um this is you know how this whole code Works basically you'll have to uh give uh initiate the client and inside the client you'll have to give the model just like a regular completion API then you'll have to give the prompt you can give the you know size of the image that you want the quality U you know whether you want HD or standard that should be fine and N would be one and it would probably remain one because you know with di 3 you can only have one generation but you can always change this n to be you know um like you know somewhere 1 to 10 in between uh for Del to for to get more um obviously images right and obviously you know the more images you create the more you'll be charged so if you increase the you know n then obviously you know you'll be charge accordingly so nals to 2 would be charging you twice right um so yeah like that so this is uh the prompt and this is the image as you can see was generated and each image can be returned as either URL or it can be uh returned in uh you know base 64 data so it would be a um I believe it would be a vector uh sorry it would be a a matrix of vector uh yeah multiple vectors in it and yeah URL will expire after one hour as well right uh Del to uh the edits so the edits that you can do in those images do you have to in this case uh you have to provide an extra you know thing in here which is uh the mask and this would both of these has to be in PNG format that's another problem um so PNG would you know make sure that you know you it shows the image what part is transparent right and then you give the prompt basically and then it does the rest of the part um so yeah as they say it should be PNG Square uh and it should be 4 MB in size um yeah less than 4 MB right uh they both must have the same Dimension so if uh you know your mask is has a lower dimensions then you know it would obviously not work uh non-transparent areas of mask are not going to be used so basically the area that is shown here that is that you have marked as transparent only this area would be used the rest of the area would not be used at all uh so they don't necessarily need to match the original image like the example about okay uh variations so basically in this case you take an image you you don't give any prompt to anything as you can see you just leave it like this and it's going to basically you know create variations of that image and you can you know create as many variations you want Nal to 1 in that case you know you'll have a one output obviously uh when you increase the output then you know it's going to increase right okay language specific uh tips so there as you can see it has two nodejs and python but since you know this code this whole code is python based so there are just few you know language specific things first one as you can see using inmemory image data so basically you don't have to import an image like this um you don't have to yeah I believe this is what they have been doing with other yeah so you don't have to import an image in this way you can always uh do it in this way in which if you have your image you know created in as a matrix then you can use that in here and you can convert that to BYO and basically you can give the that whole bite array right in here so yeah you can give that and in that case you know it it's it's just going to work the same way right um operating on image data so yeah this is another thing so basically um you could use this uh Library if you want and you can have changes that you can do into the image before you give that image to you know for for like actual like you know before you give that image to the API right so as you can see they have imported and they opened this image then they resized the image and then they went ahead and basically created that into by stream and they give that um to the API to work right now there are going to be cases in which let's say if your prompt has some issue or if you have given um you know I don't know some some some sort of things that are um like you know not a good thing in general like you know they could be um I don't know some sexually related thing or like those kind of things in general uh some Politically Incorrect maybe I would say um so in that case you are obviously you know your model might error out um and in that case you can always explore the error with this thing so basically you can take that response uh open open error E and then yeah you can explore you know what was the error uh in the request that you did so that's the the case obviously you can also explore you know the the response that you get but yeah you can explore the errors as well so yeah this is how it's done so yeah with this we come to the end of the theory part the Imaging like imagination part is really simple it's not as big as you know fine-tuning or bad ofb and stuff reason being we don't have much here uh we will just start with code in the next tutorial and uh yeah then we'll have a project on my UD course thank you so much for watching have a good one