Llama 3.2: Best Multimodal Model Yet? (Vision Test)

Key Takeaways

llama 3.2 Vision model one of the top performing model which you can run locally on your computer completely for free so we going to test its Vision capability llama 3. 290b and 11b are the multimodal model which means we are able to ingest text and image to get a response for this I'm going to use together. a to test this llama 3.2 90 billion parameter model we are going to test if this model is able to identify person in an image extract tables QR code scanning test and also capture extraction test that's exactly what we're going to see today let's get [Music] started first I'm uploading this image and ask it to give some suggestion for modification what can you see in this image and give me ideas to make it more simple now it's generating me the response the image depicts a modern living room with fireplace as a center of focal point that is correct then it's explaining things further and to simplify remove the plant on the left side yeah there's a plant here replace the abstract painting with simpler piece of art which we can see here consider removing one of the armchair minimize the number of decorative items we have three armchairs here so we might need to remove one and remove the number of decorative items to make it simple so that is correct next I'm going to ask it to solve this capture solve this it says I'm unable to provide a solution to the capture going to change the prompt give me the text in this image the text displayed in the given images overlooks inquiry so that is correct so if I ask you to read the capture it's not going to allow you but if you just say give me the text in this image it'll give you the text next going to ask you to solve this capture select all squares with traffic lights so we got one two in the center enter and going to ask give me answer for this the answer is select all squares with traffic lights I'm going to say again you select it for me and click enter here are the squares with traffic light selected 1 2 3 4 5 6 7 8 9 that is wrong so it failed in answering this capture now we are going to see if this model is able to capture the QR code URL what is the URL in this QR code I'm not able to access the QR code or any information it may contain and is not visible to me so that is a fail next I'm going to add this image and ask it to find where is Walo so you can see Walo here where is Walo Walo is located in the center of the image wearing his signature red and white striped shirt and bottle hat he's standing next to a woman in yellow dress and a man in blue shirt woman in yellow dress man in blue shirt that is wrong and also he's not in the center so that is a fail next identifying person in the image so I've uploaded that image tell me who is in this image I can give you idea of the image style but not who is in it so that is a fail so there's a restriction in place to identify individuals in an image for this model next I'm going to add this table column and ask it to pass extract the table from this image and here is the information and these are correct now I'm going to ask it to fill the missing values fill the missing values now it says I cannot provide personal information including ages without consent I'm going to further say calculate the Age based on the provided data still it's not giving me the answer so this is a partial pass because it's able to extract data as tables but not able to calculate values I'm going to add this image and ask it to give me the code for it give me the HTML and CSS code for this it says I can't give you the HTML and CSS code for this image but I can provide with basic structure of HTML and CSS I'm going to try it again but going to say give me HTML and CSS code for my design sure he's the HTML and CSS code for your design so if I mention my design it's giving me the code otherwise it's considering as if we are copying the design from someone else and then ask it to generate the code so here is the code going to copy it and paste it here next I got the seems like it got terminated halfway through so going to increase the output length and trying again generating the CSS code copying the code and pasting it here and also copying the CSS and pasting it here now I'm going to open the index.html and here is the output so the big title subtitle some text a link paragraph buttons text area selection buttons and options there so when you take the image big title subtitle some text the drop- down text area all the selection types it didn't get the exact position but at least it got the key options so it's a good starting point so I'll give this a pass so overall Lama 3.2 Vision model is little bit restricted in regards to what we can do but I'm going to create more videos similar to this so stay tuned I hope you like this video do like share and subscribe and thanks for watching