Wav2Lip: Make ANYONE speak ANYTHING (Easy Deepfake) | Online (Google Colab) Cominlooks 2020-12-17 | Making ANYTHING SPEAK ANYTHING (EASY DEEP FAKE)Hi, today we are making anyone speak anything.By, basically lip syncing, we will be making deep fakes, so you can expect to see some talking memes on the way.Colab Link (Wav2Lip): PLEASE CONCIDER TO SUBSCRIBEcolab.research.google.com/drive/14FKWR-EWYkTPFZKT8n6P-RmsIq8dny8g?usp=sharingColab Link (LipGAN): colab.research.google.com/github/tugstugi/dl-colab-notebooks/blob/master/notebooks/LipGAN.ipynbPaper: arxiv.org/abs/2008.10010We’ll be using a tool named wav2lip on a google colab notebook, which means you’ll be able to follow along, with nothing but the browser you are currently using. So no excuses are allowed.All this technique needs is a short video and an audio speech, and the speech determines what the person on the video speaks out.That's it and that's the idea.Before I get your hands all dirty, hold your horses, to appreciate the work of authors and contributors who made this.Following the good practise, make a copy of this script to your drive and make sure the runtime is on GPU.You would like to execute the first cell. This setups wav2lip by taking care of all the code, its dependencies and the pretrained models.This will take some time. Mine took about 5 minutes to complete.Now, this colab script can make lip sync deep fakes from any youtube video or from the video files you have. The second cell takes the url for any youtube video.I’ll paste a youtube url in this field.You need to make sure every frame of the video has a face on it. It's unlikely that every frame of a youtube video will have a face on it, so you can trim the video starting at some second and ending at some second. Execute this cell to download that youtube video to colab. When it's done with that, you can preview the trimmed video.The third cell takes your audio. You can either upload an audio file or record your voice now.I am going for the record option. The moment you execute this cell, it will start to record and will need to click a button that will appear to stop it.Here we go.Finally you get to preview your recordings.In the last cell, you can ignore or leave the settings for the most part and just execute this cell, to get cracking the deep fake.Meanwhile, you want to note the length of the trimmed video and your audio. I recommend the length of the trimmed video is equal or longer than audio length.Also there is a clever trick to get the fastest GPU provided by colab, this way you get the result fast, just check the youtube cards above now to get to that video.After it's done, you can preview the deep fake right here. If you are willing to download the video, you can find it in this particular path. Once found, you can download it.Here is a tip for you, notice the discontinuity between the lips and chin of donald trump. The fix is very simple. All I need to do is add some bottom padding. I will set the ‘pad bottom’ value to 30 and re-execute the cell. That fixed the issue.And finally, expand the last section, which takes the video you upload instead from youtube. Upload your video to colab from your local computer, copy the path to this video and paste it right here. When it comes to trimming the video, if you don't want to change anything, just use -1 in both the start and end fields. Finally execute it.When it comes to audio, it is identical, you have two options. This time I am selecting the upload option. You might want to execute this cell and then upload your audio file.Similarly, the last cell is also identical. Just execute this to process the lip sync video. As this is on the progress, let me add some more tips for better results.After all the changes in the setting you need to re execute the last cell again.With the help of wav2lip you can now sync lips to any dubbed movie, lecture, or conference as if the person was himself or herself speaking the foreign language.Here is a comparison of mona lisa lip synced. On the left is the result from last year's state of the art LipGAN and on the right is the result of Wav2Lip. It is obvious, wav2lip outperformed and is significantly better.In case anyone wants to play with LipGAN, you can find the colab link in the description.Technology is getting better and it's getting hard to differentiate what's real and fake.with great power comes great responsibility.we don’t hold any responsibility for any wrongdoings from the tools learnt from this video.If you are curious how wav2lip works there is the link to the paper in the description.Please like, share if you enjoyed this video and you want to subscribe and ring that bell for exciting videos to come like making 3d models from just the portrait or ai predicting someone’s face just from their voice all in google colab.#AI #Deepfake #Easy #howto #tutorial #athome #meme #memeGenerator #hacker #cloud #google #colab #free #5min #audioBasedDeepfake #technology #online #comparision #LipGan
Leetcode 24 Intersection of Two Linked Lists Cominlooks 2025-02-08 | Given the heads of two singly linked-lists headA and headB, return the node at which the two lists intersect. If the two linked lists have no intersection at all, return null.
Leetcode 23 Binary Tree Postorder Traversal Cominlooks 2025-01-04 | Given the root of a binary tree, return the postorder traversal of its nodes' values.
Leetcode 22 Binary Tree Preorder Traversal Cominlooks 2025-01-04 | Given the root of a binary tree, return the preorder traversal of its nodes' values.
Leetcode 21 Single Number Cominlooks 2025-01-04 | Given a non-empty array of integers nums, every element appears twice except for one. Find that single one.You must implement a solution with a linear runtime complexity and use only constant extra space.
Leetcode 20 Linked List Cycle Cominlooks 2025-01-04 | Given head, the head of a linked list, determine if the linked list has a cycle in it.There is a cycle in a linked list if there is some node in the list that can be reached again by continuously following the next pointer. Internally, pos is used to denote the index of the node that tail's next pointer is connected to. Note that pos is not passed as a parameter.Return true if there is a cycle in the linked list. Otherwise, return false.
Leetcode 19 Valid Palindrome Cominlooks 2025-01-04 | A phrase is a palindrome if, after converting all uppercase letters into lowercase letters and removing all non-alphanumeric characters, it reads the same forward and backward. Alphanumeric characters include letters and numbers.Given a string s, return true if it is a palindrome, or false otherwise.
Leetcode 18 Best Time to Buy and Sell Stock Cominlooks 2024-12-26 | You are given an array prices where prices[i] is the price of a given stock on the ith day.You want to maximize your profit by choosing a single day to buy one stock and choosing a different day in the future to sell that stock.Return the maximum profit you can achieve from this transaction. If you cannot achieve any profit, return 0.
Leetcode 17 Pascals Triangle || Cominlooks 2024-12-26 | Given an integer rowIndex, return the rowIndexth (0-indexed) row of the Pascal's triangle.Follow up: Could you optimize your algorithm to use only O(rowIndex) extra space?
Leetcode 16 Pascals Triangle Cominlooks 2024-12-26 | Given an integer numRows, return the first numRows of Pascal's triangle.
Leetcode 15 Path Sum Cominlooks 2024-12-26 | Given the root of a binary tree and an integer targetSum, return true if the tree has a root-to-leaf path such that adding up all the values along the path equals targetSum.A leaf is a node with no children.
Leetcode 14 Minimum Depth of Binary Tree Cominlooks 2024-12-26 | Given a binary tree, find its minimum depth.The minimum depth is the number of nodes along the shortest path from the root node down to the nearest leaf node.Note: A leaf is a node with no children.
Leetcode 13 Balanced Binary Tree Cominlooks 2024-12-25 | Given a binary tree, determine if it is height-balanced.
Leetcode 12 Convert Sorted Array to Binary Search Tree Cominlooks 2024-12-25 | Given an integer array nums where the elements are sorted in ascending order, convert it to a height-balanced binary search tree.
Leetcode 11 Maximum Depth of Binary Tree Cominlooks 2024-12-21 | Given the root of a binary tree, return its maximum depth.A binary tree's maximum depth is the number of nodes along the longest path from the root node down to the farthest leaf node.
Leetcode 10 Symmetric Tree Cominlooks 2024-12-21 | Given the root of a binary tree, check whether it is a mirror of itself (i.e., symmetric around its center).
Leetcode 9 Same Tree Cominlooks 2024-12-21 | Given the roots of two binary trees p and q, write a function to check if they are the same or not.Two binary trees are considered the same if they are structurally identical, and the nodes have the same value.
Leetcode 8 Merge Sorted Array Cominlooks 2024-12-21 | You are given two integer arrays nums1 and nums2, sorted in non-decreasing order, and two integers m and n, representing the number of elements in nums1 and nums2 respectively.Merge nums1 and nums2 into a single array sorted in non-decreasing order.The final sorted array should not be returned by the function, but instead be stored inside the array nums1. To accommodate this, nums1 has a length of m + n, where the first m elements denote the elements that should be merged, and the last n elements are set to 0 and should be ignored. nums2 has a length of n.
Leetcode 7 Remove Duplicates from Sorted List Cominlooks 2024-12-21 | Given the head of a sorted linked list, delete all duplicates such that each element appears only once. Return the linked list sorted as well.
Leetcode 6 Climbing Stairs Cominlooks 2024-12-21 | You are climbing a staircase. It takes n steps to reach the top.Each time you can either climb 1 or 2 steps. In how many distinct ways can you climb to the top?
Leetcode 5 Sqrt(x) Cominlooks 2024-12-20 | Given a non-negative integer x, return the square root of x rounded down to the nearest integer. The returned integer should be non-negative as well.You must not use any built-in exponent function or operator.For example, do not use pow(x, 0.5) in c++ or x ** 0.5 in python.
Leetcode 4 Search Insert Position Cominlooks 2024-12-20 | Given a sorted array of distinct integers and a target value, return the index if the target is found. If not, return the index where it would be if it were inserted in order.You must write an algorithm with O(log n) runtime complexity.
Leetcode 3 Find the Index of the First Occurrence in a String Cominlooks 2024-12-19 | Given two strings needle and haystack, return the index of the first occurrence of needle in haystack, or -1 if needle is not part of haystack.
Leetcode 2 Remove Element Cominlooks 2024-12-18 | Given an integer array nums and an integer val, remove all occurrences of val in nums in-place. The order of the elements may be changed. Then return the number of elements in nums which are not equal to val.Consider the number of elements in nums which are not equal to val be k, to get accepted, you need to do the following things:Change the array nums such that the first k elements of nums contain the elements which are not equal to val. The remaining elements of nums are not important as well as the size of nums.Return k.
Leetcode 1 Remove Duplicates from Sorted Array Cominlooks 2024-12-17 | Remove Duplicates from Sorted ArrayGiven an integer array nums sorted in non-decreasing order, remove the duplicates in-place such that each unique element appears only once. The relative order of the elements should be kept the same. Then return the number of unique elements in nums.Consider the number of unique elements of nums to be k, to get accepted, you need to do the following things:Change the array nums such that the first k elements of nums contain the unique elements in the order they were present in nums initially. The remaining elements of nums are not important as well as the size of nums.Return k.
GLIDE: Easiest Conditional Inpainting in Colab tutorial Cominlooks 2022-05-01 | A simplified and impressive notebook that can places objects in the highlighted portion of the existing image maintaining the style, quality, and consistency with GLIDE that is based on Diffusion models that outperform GANS models.colab.research.google.com/drive/1pr1sESu9h7rRZ8vRibLXreapak_pR1Cc?usp=sharingHi everyone. Today I'll share you a colab notebook that can paint out a portion of the image and replace them based on the text description you provide. I could take Putin, highlight the area under his nose, then provide a description, like a man with Hitler's mustache. The model should generate something like this. DALLE 2 has just blown the internet with its astonishing results creating realistic images and art from natural language. They have also added image inpainting just like what we are about to do.DALLE 2 hasn’t been public yet, when it does we’ll be making video about it as well, so stay connected by subscribing to the channel. For now, GLIDE is the best we have so before we get hands-on with the notebook let's take a moment to thank the authors and contributors. If you visit the paper for GLIDE, you will come across many examples that can give the idea of what's possible with the model. You can change the color of your hair, and replace objects and animals. What surprises me is how well the added object fits the lighting, reflections, and shadows. ### Guiding the colab notebookLet's get back to the notebook itself, you can grab the link to this colab notebook in the description, which is simplified for your convenience.Run the first cell to configure GLIDE on the colab. This should get the dependencies, clone the code and download the pre-trained model.After about 2 minutes, the cell should be executed.In the next cell, it needs your input image. You can either get the image url and paste it in the field before you execute them or leave the field empty; in such case, you can upload the input image from pc.After you execute the cell, you can easily highlight the image area, then enter your text prompt. You can then click finish, which will start the inpainting process. Finally, after a few minutes you will get the output. GLIDE has fine-tuned its model for specifically the inpainting task and it uses the diffusion model, so does DALLE 2 to generate the images you see instead of using GANS.If you want to know a few things about the diffusion model you can head to the previous video we made about GLIDE to a specific timestamp with the above link now.I am trying to monetize my channel and be able to pay for a cup of coffee, so I am coming up with many existing videos like inpainting, audio-driven facial animation, styleGAN, styleCLIP, and many more all in simplified colab notebooks. So I would very much appreciate your support, please spread this video like wildfire, hit the like button, subscribe, and ring that bell.You could also benj our previous videos like deepface for facal expression, cloning someone's voice, upscaling tiny images, coloring old black and white images, making slow mo videos from regular video files and many more.Have a great day.Thank you.
GLIDE: Easiest Text to Image with Diffusion model (colab) Cominlooks 2022-04-02 | A simplified notebook that takes a short text and imagines photo-realistic image output with GLIDE (state-of-the-art cutting edge model) that is based on Diffusion models that outperform GANS models.Colab: colab.research.google.com/drive/1pn1pP-IrM0BKvSXVeeEB_2akDz_jYV4S?usp=sharingHi everyone. Today we'll be looking at the colab notebook that takes in a textual promt and the notebook generates a photo-realistic image based on the text.Impressive. We could try something like a hot girl. Oh no. I think it's going for literally a hot dog instead. Before I guide you through the colab notebook, learn about how this kind of image synthesis is possible with diffusion models, how and with what data was it trained on, limitations and its competitions like dall.e, let's take a moment to thank to the authors and contributors for GLIDE.If you visit the paper for GLIDE, you will come across many examples that can give an idea of what is possible with the model. These are the text promt given and the image generated by the model. We have einstein dressed in a superman costume and painting of some psychedelic hamster. What surprises me is the variety of style the model can generate from photo-realistic images, oil painting, pixel art, crayon drawing, and many more.### Guiding the colab notebookLet's get back to the notebook itself, you can grab the link to this colab notebook in the description, which is simplified for your convenience.Run the first cell to configure GLIDE on the colab. This should get the dependencies, clone the code and download the pre-trained model.After about 2 minutes, the cell should be executed.Colab script: colab.research.google.com/drive/1pn1pP-IrM0BKvSXVeeEB_2akDz_jYV4S?usp=sharingIn the next cell you need to type your prompt and execute the cell and that's it. You will first see a 64 by 64-pixel image, then it is upscaled to 256 by 256 pixels.If you don't like the image you just created and want to add a few things in the image at particular spots to match more complex prompts, don’t you worry comrades. Next week we will be looking at conditional inpainting with GLIDE which is capable of making realistic edits to an existing image based on the text prompt. This allows you to iteratively improve and edit your output by painting out a portion of an image and replacing or adding objects and creatures based on the text prompt. So you better subscribe now.### Diffusion modelIn many generative applications like image synthesis, we can see GANS being used extensively, but GLIDE is based on a diffusion model which is known to generate superior image quality.When we train a diffusion model, we have an image dataset and for each image in the dataset we go through n steps adding a small amount of random noise every time. After infinitely many steps the output becomes complete noise with normal distribution like this.Since on each successive step only small noise is added, hence the idea is to train a model that takes in some intermediate image and predicts the previous step image that is slightly sharper. If we repeat this reversed process to predict the previous sharper image starting from random noise almost infinitely many times, we can generate some photorealistic image given only noise. This is the basis of the diffusion model.Note that in each step the model is predicting the noise that needs to be reduced to get the previous step image.### Guided Diffusion ModelBut that's not all, we can guide our diffusion model to generate a certain kind of image. Let’s start again with the image dataset but given the class labels for each image this time. The model would be trained to predict the previous sharper image given the image with noise, the time step and the class label of the image. For example in this case the image label would be cat.This would be the general architecture of GLIDE. We have a text network that takes the text prompt and we have our diffusion network which takes the noisy image and text features from the text network and the result id the sharper image.I am trying to monetize my channel and be able to pay for a cup of coffee, so I am coming up with many existing videos like inpainting, conditional inpainting, audio-driven facial animation, styleGAN, styleCLIP, and many more all in simplified colab notebooks. So I would very much appreciate your support and especially your feedback on what you like and hate so that I could come up with better approaches to presentation and topics for the videos. So if you like my videos and want them coming, please spread this video like wildfire, hit the like button, subscribe and ring that bell.You could also benj our previous videos like deepface for facal expression, cloning someone's voice, upscaling tiny images, coloring old black and white images, making slow mo videos from regular video files and many more.Video: Sajil Awale
K-Nearest Neighbor (KNN) Explained with Implementation | MNIST digit dataset Cominlooks 2022-03-04 | K Nearest Neighbor (KNN) is one of the simplest supervised ML algorithm that works surprisingly well for classification problems. We have presented about it in the video.Colab Link: colab.research.google.com/drive/1v-8sspWb9Jasli5JGRe6GDG5oXs30fPO?usp=sharingSlides Link: docs.google.com/presentation/d/1TjWCjzi_nxxxPplZ6x_7V-Wj8Ug7CMIDCz1BmOnJvxA/edit?usp=sharingHi, I am Sajil Awale,it's been a long time since i uploaded any video. Today we’ll be discussing K nearest neighbor or KNN. It is one of the simplest supervised machine learning algorithms for classification as well as regression problems that might be useful. We’ll briefly go through the algorithm, limitation and then work on the implementation code, that will classify mnist digit dataset and you will be shocked with the accuracy we can get.KNN is simply a 3 step algorithm; when you get a test input, you need to calculate the distance with all the training dataset, then find the K training set that is closest to the test set, finally we assign the label for the test set based on the majority label of the K training set. Here K is just a number. When k is equal to 3 as shown in the first picture, we take 3 nearest training sets from the test set. In this case out of 3 sets, 2 of them are of class B, hence we decide the test set to be of class B. Similarly, when k is equal to 6, out of the 6 closest train sets, 4 of them were of class A, hence we predict the test set is of class A.Obviously, one big question is how can we select the value of K. There is no hard and fast rule, we usually start with k = 1 and continue testing at a higher value of k. When we use a lower value of K, the model is low biased and high in variance, which means it does good with training sets but not good with test sets. Similarly, if we use a higher value of K, our model will be highly biased and low in variance which means it is not so good with the train set but better with the test set.This graph summarizes this, as you increase the value of k, the error on the test set decreases and starts to increase again but for the train set, error increases. The value of K where error on the test set is minimum is the optimum value of K. Since we need to store all the training dataset, the space complexity is large. Also we need to compute the distance of each test set from all the training dataset, time complexity is also big. To solve these issues, we can either only take important features and use them only for computation or we can reduce the overall training dataset. In the implementation we are purely using only python and numpy with a mnist digit dataset. Just to recall, mnist has 60 thousand training images and 10 thousand test images of 28 by 28 monochromatic pixelsAnd here is the high level understanding of our implementation. We have our mnist dataset, we took the first 6000 images for training and last 1000 images for testing. Then we calculated the euclidean distance of each test image from all the training images and sorted them in ascending order. Now for each value of k starting at 1,we predict the digit based on the digit majority of the K nearest training image. Then we find the accuracy of KNN for each value of K. You can find the implementation in the colab link attached in the description. It's a simple implementation, we also have the flow chart for each function we defined in the slides, please find the link in the description. As the final output we have an accuracy graph, on the x axis is the value of , Y axis we have accuracy and these different curves on the graph are the accuracies of each digit from 0 to 9 as we vary the value of K. For example, the yellowish dotted line is the accuracy for digit 8, which seems to be lowest among others and the orange dotted line is the accuracy for 1,which has the highest overall accuracy. Looking at this graph the optimum value of K in this case is 5, as we have the highest accuracy there.We would very appreciate your support, so please spread this video like wildfire, hit the like button, subscribe and ring that bell for exciting videos to come like about self organizing maps which is an unsupervised ML algorithm that has multiple interesting applications like clustering,classification and dimensionality reduction. We are going to discuss the algorithm with beautiful and clear visualization of each step along the way. Please comment on any ML algorithm you want to see being covered in coming videos.You could also benj our previous videos like deepface for facial expression, upscaling tiny images, coloring old black and white images, making slow mo videos from regular video files and many more.Have a great day.Thank you.
Liquid GAN: Easiest Deepfake for Motion Transfer on Google Colab (meme generator) Cominlooks 2021-05-30 | Today we are making this. Uyrr… transferring full-body motion from someone in a video to a person on a still image. And will take no more than 10 minutes.Colab: colab.research.google.com/drive/1zUgh4nYsD-AlqpPPKwBfDLDzj_CpqMj7?usp=sharingPaper: arxiv.org/abs/1909.12224?source=post_page-----e74d7d347c2We are not going through the science but straight to making one for yourself with google colab.Before I get your hands all dirty, hold your horses, to appreciate the work of authors and contributors for Liquid GAN.Grab the first link in the description, that will direct you to this colab script which is heavily simplified for convenience.You can start by making a copy of the script to your drive.Run the first cell which will set up the environment, install dependencies and download pre-trained models. Go make yourself some coffee, this will take a moment, like 5 minutes.The second cell will take all the inputs. Before you give any input though, execute the cell.There are 2 major inputs, the source image, to which the motion is transferred, and the reference video, from which we extract the motion.You can provide the source image in various ways, and they are listed in different tabs. You can either pick one from the provided sample images for demo or upload your own image or refer image straight from the web.For now, I will be using one of the sample images.In the same fashion, you can mention your reference video from the provided sample, get from a youtube video, upload your own video or refer to a video from the web.All of the provided sample videos have someone dancing in them. I will select one of the provided sample videos.The start time and end time are used to trim the video in case they are long. Setting them to -1 leaves the video untrimmed. Other inputs like pose_fc and cam_fc can be left unchanged. Finally, execute the last cell. This will take some time depending on the length of the reference video. Well, you can get back to watching cat videos on youtube until it is done.Mine took about 10 minutes.When it's done your output is displayed.That is easy as it gets.Now let's say you have a source image of the target person. It could be your friend or just some victim of your meme.It is best if the source image has only one person standing, preferably in a T pose and with the distinct feature of legs and arms.Go back to the second cell, and switch to “upload from pc” tab. Here, out of 3 upload options only the first which is for the front facing image is essential. Back facing and the background image are optional to remove artifacts.Sometimes when the person on the reference video turns around, liquid gan reuses the texture resulting in 2 face persons on the front and back of the head. Adding the back-facing image will solve this issue. To further remove any artifacts due to moving subjects you can add the clean background image without the person on it.Once done make sure you don’t change the tabs again.For reference video, you can again go with the provided sample video but let's take a youtube video. Youtube videos can be long and we might just be interested in a small portion of the video. So,just note the start and end time of the video you're interested in trimming. Similar to source image, you want the trimmed video to have only one person with distinct features of legs and arms and steady camera for best results.If you have multiple people in them you want to crop the video and upload them.Just paste the youtube Url link and the noted start and end time in seconds.You can ignore the rest and execute the last cell. After a good break, your results will be waiting for you.Looks like KK lost all her curves there. If you don't want that to happen to your model, you can have images taken at multi-angles and upload them all together as your source input. This will get you the best results.Technology is getting better and it's getting hard to differentiate what's real and fake.Remember with great power comes great responsibility.In Fact, we don’t hold any responsibility for any wrongdoings from the tools learned from this video.If you are curious how LiquidGAN works there is the link to the paper in the description.We would very much appreciate your support, so please spread this video like wildfire, hit the like button, subscribe, and ring that bell for exciting videos to come like AI predicting someone’s face just from their voice or image inpainting on a simplified notebook.You could also benj our previous videos like deep face for facial expression, cloning someone's voice, upscaling tiny images, coloring old black and white images, making slow-mo videos from regular video files, and many more.Have a great day.Thank you.#liquidGAN #deepfake #tutorial #easy #colab
RIFE: AI makes Slow motion videos in Real Time (Simplified google colab) Cominlooks 2021-05-04 | MAKING SLOW MOTION VIDEO FROM REGULAR VIDEOS WITH RIFE | AIPreviously, we have seen Super slow mo, which can take a regular video and turn it into a high fps video file with frame interpolation.But you might have noticed it was an extremely slow process.Today we are discussing R I F E, which generates slow-motion video almost in realtime and results are indistinguishable from the previous one.You can watch the previous video on slow mo by selecting the cards above now.PLEASE CONSIDER TO SUBSCRIBELink to colab script: colab.research.google.com/drive/16WbKFti_Ss4_2tjoMcpopCwdAN7MGId8?usp=sharingPaper: arxiv.org/abs/2011.06294Get FASTEST GPU video: youtu.be/uQAbyVk36xUBefore we get right in, lets’ appreciate the authors and contributors of RIFE.We’ll be making slow mo videos using google colab providing you with free gpu.So jump to the first link in the description, that will direct you to a colab script, which is simplified for your convenience.This script can slow down the videos from either youtube or the videos you upload.Following good practise, make a copy of this script to your drive, and make sure you are on a GPU runtime.Run the first cell to set up your machine with the software.The second cell will make slow mo against any youtube video.You will need to fill in the Youtube url where you can paste the link to any youtube.As for example, this is a youtube video. I will copy the url of this video and directly paste it to this field.SLOW_MOTION_FACTOR takes one of value from drop down, which describes how slow you want the video to be. Higher this value, slower or higher is the fps. For this tutorial I leave this field at 8. To run this process for the entire video may take a very long time. So you may trim the video starting at some seconds and end the clip at some seconds. For this tutorial I set the start time at 51 and end at 53 seconds. Finally, execute this cell. This will take some time as it crunches the source video to slow mo video.While that's running,check the youtube cards above now and learn the tricks to get the fastest GPU in colab, link to video tutorial is also in description. Obviously, faster the GPU, shorter is this waiting timeAfter all that's done you are left with a comparison video. The clip at the left is the output result and the clip at the right is the source video.As expected it did a decent job to interpolate between frames.You can simply download the slow mo video by executing the next cell.But a single clip cannot describe how awesome this software is. You might want to try this yourself.The next cell will take in your video file and turn it to slow mo video.First you need to upload your video to the colab machine through the file tab. After your video is uploaded, identify it with its name. Right click on it and select the copy path option. Then, paste this path to the PATH_TO_YOU_VIDEO field. In the same fashion fill all the fields like so. And execute it.After all the processing you are left with another comparison video. If you want to download just execute the next cell.Now if you are curious how RIFE works. There is a link to the paper in description. In nutshell, these algorithms are performing frame interpolation, that is adding single or multiple intermediate frames to make existing video slower. Both RIFE and super Slow mo are based on optical flow. Hence the video has to follow constraints of optical flow like the pixel intensity of a point should be constant throughout the video otherwise these approaches might generate artifacts.I hope you found this video fascinating or helpful. We would very appreciate your support, so please spread this video like wildfire, hit the like button, subscribe and ring that bell for exciting videos to come like deepface to transfer human motion, AI predicting someone’s face just from their voice or making 3d models from just a portrait.Thank you.#rife #slowmo #AI #tutorial #easy #fun
PIFuHD: AI constructs 3D model from single portrait image (google colab) Cominlooks 2021-05-02 | Hi, today we are making 3d models of people with nothing but a single portrait image. One picture and you have a complete 3d model of the person. You could rig this model, animate them, 3d print it and do all kinds of cool stuff, the sky's the limit.We’ll be using a tool named PIFulHD on a google colab notebook, which means you’ll be able to follow along, with nothing but the browser you are currently using. You don’t need to bother about the dependencies, libraries to install or set up anything on your computer. Everything is handled on the cloud, just for you.So no excuses are allowed.Note that this method works only for humans. If you want to recreate a 3d model of any object that is not human, check the cards to be directed to that video. Before I get your hands all dirty, hold your horses, to appreciate the work of authors and contributors who made this.Grab the first link in the description, that will direct you to the colab script.This is oversimplified, easy to use and made for your convenience.PLEASE CONSIDER TO SUBSCRIBEcolab.research.google.com/drive/1NzP4oI_KighbpfEEVCnYKZ0lfzDpqTyE?usp=sharingFollowing good practice, make a copy of this script to your drive and make sure the runtime is on GPU.I’m referring to the GPU on the cloud provided to you for free by google.You would like to execute the first cell. This setup PIFulHD by taking care of all the code, its dependencies, and the pre-trained models.This will take some time. Mine took about 4 minutes to complete.This colab script can take images from the web or you can upload your own image to create the 3d models. Let me demonstrate both the cases one by one.For the first case, I am going on the internet and will need to give the URL to the portrait image.Let this be the portrait. I will right-click on it, and copy the image url. Going back to the colab, I will paste the link to this field and finally execute this cell. Give it a few minutes and you have a preview of the 3d model just created. To download the 3d model, just execute the cell just below this one. Your 3d obj file should download immediately. You can view the obj file with many programs, use blender, open-source software, view, clean, modify, add texture, or animate the model as you wish.In a very similar fashion, you could make the 3d model for the image you upload.In this case, you will need to execute the cell first. You will be asked to upload the image, just select the right image from your local computer. After a few minutes, your model is ready and you can preview it in the video. Finally, to download this 3d model, obj file, just execute the next cell.In the end, I want to share very helpful tips for getting the best results from this tool.You want to use high-resolution images, low res images can create artifacts. You also want the image to have a single person standing and facing the camera. The image should be well uniformly lit; harsh shadows, very dark or bright pictures can introduce more artifacts.You want the background to be simple, you could also try using an online tool like remove.bg for removing background beforehand. The model is trained on humans, so this won’t do good with anime characters.Finally, you could search for hashtag pifuhd in twitter to find many examples for reference on what works and what fails. Here are a few examples listed. If you are curious how Pifuhd works there is the link to the paper in the description. You aren’t into papers, there is an entertaining video by 2-minute paper explaining the details.I hope you had fun.Please like, share, and comment on your suggestions, if you enjoyed this video and you want to subscribe and ring that bell for exciting videos to come like ai trying to split voice and instrumental from songs or ai predicting someone’s face just from their voice; all in google colab.Thank you#pifuhd #ai #3d #colab
MAKE: Slow motion videos with AI on cloud for free (colab) Cominlooks 2020-09-16 | MAKE slow-motion video from normal videos with AI on the cloud for free using google colab.PLEASE CONSIDER SUBSCRIBING. THANK YOU Link to COLAB script: colab.research.google.com/github/AwaleSajil/SuperSlowmoColab/blob/master/SuperSloMo_Simplified(youtube)_V2.ipynbGet FASTEST GPU video: youtu.be/uQAbyVk36xUPaper: http://jianghz.me/projects/superslomoSo you want to create slow mo videos but you don’t own any expensive slow mo cameras.You probably could slow down a normal video.And it looks choppy and bad. The more we slow it down, the more it looks like a slideshow. With the help of an AI tool named Super Slow Mo, we change all that to this smooth slow mo, which you can process on cloud GPU, online, all for no money, so you save time and your dollars. And this process is much faster than previous methods.Hi my name is Smriti. We have covered fancy topics and have interesting videos on the line, so you might want to subscribe to the channel. Before we get right in, lets’ appreciate the authors of Super SlowMo.Jump to the first link in the description, that will direct you to a colab script, which is modified to be super easy to use.This script can slow down the videos from either youtube or the videos you upload.Following good practise, make a copy of this script to your drive, and make sure you are on a GPU runtime.Run the first cell to set up your machine with the software.The second cell will run super slow mo against any youtube video.You will need to fill in the Youtube url where you can paste the link to any youtube video but I am not sure this works with music videos though. As for example, this is a youtube video. I will copy the url of this video and directly paste it to this field.SLOW_MOTION_FACTOR takes in an integer, which describes how slow you want the video to be. One means the speed of the original source video. Two means you want the slow mo to be at half speed and so on. For this tutorial I leave this field at 8. To run this process for the entire video may take a very long time. So you may trim the video starting at some seconds and end the clip at some seconds. Finally, execute this cell. This will take some time as it crunches the source video to slow mo video.While that's running,check the youtube cards above now and learn the tricks to get the fastest GPU in colab, link to video tutorial is also in description. Obviously, faster the GPU, shorter is this waiting timeAfter all that's done you are left with a comparison video. The clip at the left is the result of super slowmo and the clip at the right is the source video.As expected it did a decent job to interpolate between frames.You can simply download this video. If you want to download just the slowmo video, then scroll your file tab to find a file named output underscore slow dit mp4. Download that file.But a single clip cannot describe how awesome this software is. You might want to try this yourself.The last cell will take in your video and apply super slowmo on it.First upload your video to the colab machine through the file tab. Click the weird upload button, select the video, and hit open. After your video is uploaded, identify it with its name. Right-click on it and select the copy path option. Then, paste this path to the PATH_TO_YOU_VIDEO field. In the same fashion fill all the fields like so. And hit execute.After all the processing you are left with another comparison video. If you want to download just that slow mo video, you need to scroll for “myoutput underscore slow dot mp4”.Now if you are curious how Super slow MO works. There is a link to the paper in the description. Or if you are not that kind of person there is an entertaining video by two minute paper. In nutshell, these algorithms are performing frame interpolation, that is adding single or multiple intermediate frames to make existing video slower. This is mostly done in 2 basic ways. The first is frame blending which basically computes average between two frames. This technique rarely generates any visual artifacts but it lacks the understanding of motion that is happening. The other technique called optical flow is much smarter as it understands the kind of motion in the video, like translation and rotation and does much better in most cases but this method may take a long time to process and may introduce artifacts.The super Slow mo is based on optical flow, is much quicker to render than other competitors, and is adapted by Nvidia.Thank you.#slowmo #AI #tutorial #easy #funMusicYou’re free to use this song in any of your videos, but you must include the following in your video description:Easy Lemon by Kevin MacLeod is licensed under a Creative Commons Attribution license (creativecommons.org/licenses/by/4.0)Source: http://incompetech.com/music/royalty-free/index.html?isrc=USUAN1200076Artist: http://incompetech.com
Upscale images resolution with AI on cloud Cominlooks 2020-09-04 | Learn to change any low-resolution images to high-resolution images with an AI tool named ISR, and you don’t need any setup on your local machine because we are going to take advantage of free cloud service from google colab.Link to colab script: PLEASE CONCIDER TO SUBSCRIBEcolab.research.google.com/drive/18zKHfswFF8pOOT1X7GmViGvET0R_vQBT?usp=sharingBefore we jump in, let's appreciate the contributors of the project and authors of the papers that were implemented. github.com/idealo/image-super-resolutionTo follow along with me, get in the description to find the first link that will direct you to a colab page. colab.research.google.com/github/AwaleSajil/ISR_simplified/blob/master/ISR_simplified(youtube).ipynbThis script right here is a very simplified version of ISR or Image Super Resolution. All you need to do is mention your low resolution image, run and get the high resolution image as an output. Following the good practise, I will make a copy of this script on my google drive first.Once that is done copying, a new tab will emerge.Then run the first cell that will install, download necessary libraries, weights and create models, basically setting up the environment.After that is done, the second cell takes in your low resolution image.It can be from the web or you can upload your own image.You can search for low resolution images in google, just use the tool option and specify the size to be an icon, which makes sure the image you’ll find is no bigger than 256 by 256 pixels.Just copy the Image address, then paste it to this field.In case you're uploading your own image. Upload the image to colab. Then copy the path and paste this path to the same field.Now to get the best results you need to set the model right.For that i suggest you look at the original colab script. You will learn much more about ISR, pretrained models and where it is best used step by step.In summary, the RDN noise cancelling model is best when your low resolution image has compression artefact like this.All the models will generate twice the resolution of the image. Except the RRDN model, quadruples the resolution.It's best you experiment with all different models.Selecting a model, run this cell. And finally run the 3rd cell to view the results.On the left is your generated output and on the right is the original image that has been scaled to match the resolution of output.You can move the slider left and right to compare the result.If your slider does not respond or you want to download the generated result. Open your file tab and download the image named result.pngTo better understand ISR or if you want to train your own model visit the docs page, which is short and to the point.You will learn what makes ISR, how you can train and customize your own architecture of models. In a nutshell,ISR is made with a residual network, a special type of neural network which implements a shortcut path. Data can take the longer path through all the layers or just use a shortcut.This prevents vanishing and exploding gradient problems in very deep neural networks..Also the random letters, C, D, G, x are hyperparameters of the models. Just to mention a few, D is the number of Residual Dense Blocks, C is the number of convolutional layers stacked inside a Residual block and x is the scale by which the output is upscaled.ISR allows you to change these parameters and train with your own data, to get you your desired output. And the code to perform all that is in the docs.I hope you like this video, found it informative or helpful. Please subscribe, cause most of the viewers watching are not subscribed. Also Like share and comment, that will keep me motivated.Thank you.#Upscale #Image #Tutorial
QuantumHackathon Traffic Optimisation Demo Cominlooks 2020-08-29 | Team name: DristiTeam members:Rashik ShresthaSajil AwaleShrey Niraula
How to: Get the fastest GPU on google colab for free ? Cominlooks 2020-07-17 | This video will get you the fastest GPU in colab. Before we get it on, I am giving a quick shout-out to Sina Asadiyan for sharing this trick with me. So back in your colab, when you change your runtime to GPU session in google colab. You are provided with one of a type of GPU. research.google.com/colaboratory/faq.htmlThe GPUs available in Colab often include Nvidia K80s, T4s, P4s and P100s and there is no way to choose what type of GPU you can connect to in Colab.For that you are expected to pay for colab pro for 9.99 dollars per month which i don’t think is worth that much for a few extra features given we are provided so much already for free.This video is dedicated to get your session the fastest GPU in your colab session for free.Let’s jump right to google colab. We’ll make a new notebook. And change the runtime to a GPU.Now I want all of you to type this command in your first cell. !nvidia-smi Exclamation, nvidia, hyphen, smi, then execute this cell.This command will tell you what the model of GPU installed on your system is. If you give a look to the output you can find the model of the GPU.Right now, I have a virtual instance of Tesla K80.In the layman's language. The P100s is the fastest GPU in google colab. So if you have p100s, then you are good.But if you have some other GPU and want to change it to p100s. You need to factory reset runtime and re-run that cell until you land a p100s. There is a random chance of that happening. But given enough times, you will definitely land a p100s.I wanted to write a script that will reset the runtime until we have p100s automatically but couldn’t do that. So if you are one of the guys who was able to code that, please share it to everyone here.Now to compare the given GPUs, I searched for the appropriate benchmark tool but could not find any that was compatible with colab.Now i am not sure that p100s is the desired GPU for all of you. It depends on the application. For that, given i have no benchmark results lets compare the specs.docs.google.com/spreadsheets/d/186ECC0hSplx5A9sychyUMuZn_rC8oJoz3hsAD5M6mJs/edit?usp=sharingWe compare the four GPUs based on the parameters listed.Out of four, tesla T4s is the most recently launched. This means the gpu may have newer technology embedded or be more efficient. The processing capability or speed of any GPU is parameter of its clock speed, number of cores and number of instructions performed per cycle. With that regarded, p100s has the highest core clock rate and k80 has the lowest.Any decent GPU, can be overclocked or boosted to maximize its performance. In that regard, t4s is the most overclockable. This has to do with the fact that the transistor size is smallest at 12nm, making it the most power efficient, hence lower power loss through heat and lower thermal throttling. The more the transistor counts, the better. P100s wins this game with over 15 thousand million transistors. When it comes to memory type, GDDR6 is the latest. But HBM2 is comparable to GDDR6 but made in a different way, and they excel in different areas. The high bandwidth of HBM2 is ideal in situations such as AI and advanced computing, while GDDR6 can offer a lot of the same performance with higher power requirements, but at a lower price.I compared the speed of these gpus to render the deep fake video that I instructed in the previous video. Please do check that.You can clearly see k80 performs at about 8 iterations per seconds and takes about 50 seconds to complete a render.Similarly, p100 performs at about 29 or 30 iterations a seconds. That is more than 3 times the speed performance of k80.Again, T4 performs very close to p100, that is at about 28 iterations a second. And both the p100 and t4 completes a render for about 15 seconds. This is quite impressive, given that t4 operates at less than one third the power of p100. And finally, to my surprise, I was unable to land a Tesla P4, so I cannot confirm anything about it. If this video was helpful or informative to you hit that like button, share and comment You might want to subscribe because there are a lot of interesting videos on the line. Like making slow mo videos with AI, decodifying images and videos and upscaling image resolution, all in colab. Thank you#fastest #gpu #colab #free #googlecolab #machinelearning #deeplearning #neuralnetwork #technique #tutorial #easy #computation #savetime #googlecolab #tricks #benchmark
Tutorial: AI Colorize Black and White with Realism(on Colab) Cominlooks 2020-07-11 | Link to colab Script:colab.research.google.com/github/AwaleSajil/Deoldify_image_colab_modified/blob/master/ImageColorizerColabStable_modified(youtube).ipynbLink to github pagegithub.com/jantic/DeOldifyBefore we start, let's give some appreciation to the guys who made the software called deoldify.Jump to the description of the video and get to the first link that will direct to this colab script.First things first, make a copy of this script to your drive and make sure the session is on GPU for faster processing, like so.Then run this cell to setup deoldify on your virtual computer. You’ll need to provide a black in white image. You can do that in two ways. You can either get the url of an image from the internet or upload your own image and deoldify that. To deoldify any web image. Go to the web searching for a black and white picture. Come back to your colab, and paste it to the source_url field. Not worrying about anything, you can execute this cell.You will get the colorized version of the image below.In the same fashion, to colorize your image. Upload the image through the file tab. Paste this path to the path field. Also you need to omit and make the source_url empty. Finally, execute this cell again. And there is your deoldified pic.Coming back to the settings, you have render_factor. The default 35 is great, works with most images. The lower value of render_factor renders faster, gives vibrant images but the color fills are blocky. So if you're dealing with old low res pics you can go with lower values. Higher render factors are often better for higher quality images, but the colors may get slightly washed out.If you want to experiment with the render factor, run the next cell that gives you a preview of the outputs at all possible render factors.One good use of this software.Artists usually sketch portraits, landscapes, and pretty much draw anything. This app can be used for colouring the Hyper-realistic art in realtime. I’m sure this gives them a new dimension. youtu.be/J0VF0VudUEwFlawsDeoldify isn’t perfect. It’s just a machine trying to fool your eyes.To demonstrate loss of vibrancy, I used a color vibrant painting , changed it to black and white image and fed it to deoldify.Again, tried it with a rainbow image and the rainbow faded to clouds.Small details - Left uncoloredFrom a far glance the colorized picture looks great. At a close inspection, you see her hands, the candles are completely colorless. Mismatched colorsTo illustrate the mismatched color issue, this image of the deoldified crowd has a bottle and the color of its cap is clearly inspired from the color of the face. This usually happens to objects holding a small screen area.WorkingDespite Deoldify being open source, famous on the internet, a paper explaining its working is still missing. But there is a good article from fast.ai about the technology used in it. Link in the description. fast.ai/2019/05/03/decrappifyBut summarized in brief.Deoldify initially used GANS. GANS are neural network models with two distinct pieces. A generator network that generates images from noise and a discriminator network that critiques and classifies the samples as real or fake. Together, the network generates realistic imagery.In case of Deoldify, they used a generator based on U-net, a resnet34 pre trained on imagenet and discriminator from DC-GAN.Data were constructed from the color images of imagenet which were transformed to black and white at various level of contract But GANs are notoriously difficult to train, due to requiring a large amount of data, needing many GPUs and a lot of time to train, and being highly sensitive to minor hyperparameter changes. Plus, deoldify evidently generated artifacts at the time.Later, deoldify switched to Self-Attention GANs which focused on having maximal continuity, consistency, and completeness in colorization.Finally and currently deoldify use no GAN technique. They are new and exciting techniques in GAN training in pursuit of higher quality and more stable renders.And the results are speaking for themself.youtu.be/l3UXXid04YsThe process of deoldifying does’t limit to just images but also videos. For that stay tuned, subscribe, like, share and comment, motivating me to make more high quality videos.Thank youIntended Force by Kevin MacLeod is licensed under a Creative Commons Attribution license (creativecommons.org/licenses/by/4.0)Source: http://incompetech.com/music/royalty-free/index.html?isrc=USUAN1100195Artist: http://incompetech.com#deoldify #colorize, #photo, #colab, #decrappification, #googlecolab, #deeplearning, #transferlearning, #machinelearning, #GAN,#NOGAN, #Generative #adversarial #network,#photography, #photo, #image, #picture, #how #deoldify #works, #online #tool #athome
AI Clones Voices after 5 seconds of listening (online/colab tool) Cominlooks 2020-07-02 | USE FREE CLOUD SERVICE, and CLONE ANY VOICE in no timeLink to colab scriptPLEASE CONSIDER TO SUBSCRIBEcolab.research.google.com/drive/1inbK8Py9vzhFNVq9uV3tojBKT8lTwl5C?usp=sharingIn this video we are going to make some voice clones with an online tool that you could follow along. All this new technique needs is 5 seconds; 5 seconds of someone speaking before we can clone them.Before we jump right at it. Let's take a second to thank the contributor of this software and colab script as well as the author of the papers that were implemented.Grab the first link in the description, that will direct you to a colab script.colab.research.google.com/github/AwaleSajil/VoiceClone/blob/master/RealTimeVoiceCloning%28multi_dl%29.ipynb(Has only one download link, use above link if this colab is not working)Following the good practise, make a copy of this script to your drive and make sure the runtime is on GPU. But at the time of making this video there seem to be some errors with GPU due to some recent updates, so i will be selecting None. The first cell here, sets up your colab with a voice cloning software. Lets execute this. As this cell proceeds, it downloads the actual code, the pre trained model and initializes the tool.Once that's done. The second cell is here to get your voice input. You can either record your voice right at this moment or upload an audio file of someone speaking. Let's go with the record option. Then you have a choice of how long you want it to record.Remember, the more data you feed your model, better will be the estimated output. Cause neural net models are data hungry. What fascinates me is this model can give decent results with just about 5 second of data. Let's go with 10 seconds. Then hit execute. You might need to allow your browser for the permission to use the microphone.Finally whenever you are ready to speak hit the Record Button.After the recording you can replay your voice.The final cell will synthesize or clone your voice. You are given a default text.I am overwriting the text to my choice. You can type anything you want in this cell. After it’s done cloning, you can play or download the cloned voice clip. Are you amused or creeped ? Let’s go with some non english text this time.Pretty close.What about changing the input recording to some non english speaking ?That is certainly impressive but I have to criticize the synthesized voice for lacking general human naturalness, sometimes the pauses are too long.Just for the purpose of curiosity, let’s compare my fake or synthesized voice with my original voice in audacity.I imported both the sound clips. The audio track on the top is my actual voice and the other one is synthesized. The waveform appears close. The peaks and valleys are adjacent. What about spectrograms ?In the spectrograms, frequencies are on vertical axis and time on the horizontal axis. And the color represents the energy of the voice at that particular frequency and time. The spectrograms of both fake and real are comparable. I would say the model is in its mediocre level but with few more upgrades it is up there.For learning the workflow there is a great video by two minute paper (youtube.com/watch?v=0sR1rU3gLzQ). Link in the description.But basically the workflow has 3 components. Speaker Encoder which takes your voice as input and generates a vector embedding that represents the essence or identity of your voice characteristics in compressed numerical form. The speaker encoder is trained with thousands of voices to be accurate. Next is the Synthesizer, which takes in the text, you want the subject to speak. The synthesizer with the help of voice embedding generates a synthesized spectrum of the words you want your subject to speak.The last one is neural vocoder which converts the given mel spectrum from the synthesizer to the time domain waveform.I hope the video was entertaining, useful or informative. You should know there are lots more exciting videos on the line so I want you to like, share and subscribe for more videos. Thank you.#voice #clone #AI #cloud #machine learning #deep learning #deepfake #voice #athome
Easiest DEEPFAKE Tutorial with a single image (colab/cloud) Cominlooks 2020-06-17 | Use FREE cloud service and make your own deepfake video in no time.Let's get straight, this video doesn't go on the science of how and why. But it is all about making your deep fake videos all on the cloud so you don't need to set up on your computer with gpus. If you're into the science of deep fakes, you should await our future videos.Before I get your hands dirty, I would like to and you should too be thankful to the authors of the recently published paper entitled First Order Motion Model for Image Animation which made it all possible. The recent innovations made it possible to make these face swap as well as expression swap videos with just a single image. Go to the first link in the description, which directs you to a colab script. colab.research.google.com/drive/1-egRJApHE9YoEz2ohPZBGKeKm6I-jVWv?usp=sharingIt's a good practice to make a copy of this script to your drive. Make sure you set to a GPU runtime otherwise it wouldn’t work. First things first, we configure the setup with deepfake. Select one of the checkpoints. Anyone will do, but you do get better results with vox advanced checkpoint. Then run the cell.The second cell will run the deep fake with provided sample data. There are two terms you need to know.Source image and driving video. The facial expression and features from the driving video are animated into the source image.There are few options to choose. For the source image there is a statue, some GOT characters, doll and shrek. I will choose (Mona lisa) for the tutorial.And for driving video, you have few more options. All these videos have them speaking. I will choose (Leonardo decapreo) for this tutorial.Then you execute to see results.Give it a minute, and your deep fake video is ready.Obviously, the first image is the source image, then driving video. The third is the animated source image with facial expression from driving video. And the last one has both the facial expression and features extracted to the source image.That was cool. But it's no fun, until you make them of your friends, make them look silly and roast them. The third cell is just for that.Few tips before we upload data of our own.The source image needs to be cropped focusing on just the face. After cropping, the image should be close to square, meaning the height and width of the image are nearly equal, otherwise it is subject to being stretched. The same rule applies to driving video. You will need to crop the video. I think the easiest way to do that is by using some screen recording tool to focus on the face and make sure that cropped video is close to being square.This setup supports most common videos encoded in h264.And note all the images and videos are resized to 256 pixels by 256 pixels, so don't bother uploading large resolution files. Getting a source image of someone is not difficult. For the driving video you can record yourself or get someone from the internet.In case you're not so rich with data, you could use the sample data from the drop drown.After acquiring data, you will need to upload it. Click on the file tab to your left. There is your upload button. I click it, select my data then click open. Now the data is on the progress of being uploaded.Once the upload is done, identify the source image with its file name or you can also preview it by double click. Hovering the file, right click and among the options go for the copy path.Now, at the source image field paste the path.Very similarly, copy the path to the driving video, and paste it to this field.Finally execute this cell. Give it a minute and boom you have made your first deepfake video.You can download this video with option at the right of the video player.In case the video does not playback, check your file tab. Your deep fake videos are saved in the file with the name yourDeepfake1.mp4 and yourDeepfake2.mp4. Just download these videos. The message behind this is simple. Anyone can make these fake videos, upload them to the internet claiming to be real. Remember I and this channel doesn't hold any responsibility of any wrong doings from the tools learnt from this video.In the coming video, we go into making one of those high quality deep fakes that are too convincing that requires tons of data and it's all to be done on colab. So subscribe, like and comment so i can stay motivated to bring more videos.Thank you.#easy #deepfake #meme #memeGenerator #hacker #cloud #google #colab #free #5min #single image #technology #online #plug and play
How to: Prevent google colab from runtime disconnects ? Cominlooks 2020-05-29 | Code Snippet:function ClickConnect(){console.log("Working"); document.querySelector("colab-toolbar-button").click() }setInterval(ClickConnect,60000)Medium Post link: medium.com/@shivamrawat_756/how-to-prevent-google-colab-from-disconnecting-717b88a128c0Ever been in a situation when your google colab session disconnects during a process. If you ever used google colab, it is an awesome service provided by google for training neural nets, but can also be used for other process intensive tasks.Check more of my how to type videos where I show you how you can render blender animations, 3d photo scan or even make deep fake videos utilizing generous GPU and RAM provided by google.You probably clicked this video to prevent the usual session disconnect issue. Google colab was intended for interactive purposes, so on every 90 minute of your inactivity, the session disconnects even if the process is on progress. You could solve this issue very easily. I found the fix first in this medium post, link is in the description. Thank you. All you need to do is copy a code snippet.The first code snippet used to work before, but these days gives you errors. The second code snippet still works today. Copy this code, you can find this code in the description as well.Then goto your colab tab. Reach to the 3 dot icon menu in your browser, particularly google chrome. Hover your mouse at the more tools option and finally click on developer tools. Or use the shortcut Control + Shift + I. Then reach to the console tab and finally you want to paste the code at this location and press enter.This code basically auto clicks this connect button minute, hence registering that you are interactive to the colab session and hence prevents the frequent and annoying disconnects. Thank you for watching this view. You might want to hit the like button, subscribe for more video. VoiceOver: Shrey Niraula Video: Sajil Awale
Coronavirus Active cases | 3d graph visualization Cominlooks 2020-05-06 | The 2019–20 coronavirus pandemic is an ongoing pandemic of coronavirus disease 2019 (COVID-19) caused by severe acute respiratory syndrome coronavirus 2 (SARS‑CoV‑2).This 3d graph visualization presents the curve of various countries for conformed corona active cases.Hope you are healthy and enjoyed this video. Please like, share, subscribe to this channel showing your support for upcoming videos, because it does take effort in making these videos. Code/Script:https://github.com/AwaleSajil/Covid-1...Datasource:https://data.humdata.org/dataset/nove..."Art of Silence - by Uniq" is under a Creative CommoMusic:ns license (Creative Commons – International Recognition 4.0 – CC BY 4.0)
How to: 3d photo scan free on cloud ? Cominlooks 2020-05-03 | Link to google colab script: colab.research.google.com/drive/1DR9uSfWdWWyeMxvBMvWnYA3UXIfzWyf2Imagine a physical object, it can be anything, a mug, a plant pot, a statue, a monument, a building, a hill or a mountain or even human face. ,Hi, my name is Sajil Awale and this video is all about 3D model reconstruction from 2d images, all processed on the cloud for free with google colab.You do not need to worry about not having a GPU, because you are provided with a free cloud GPU on colab.This means with a few pictures from your phone and a couple of clicks, you can reconstruct the 3d model of the object with textures on them, which you could relight on your 3d render or 3d print for replica.So to achieve this I have laid out 7 simple steps.Step 1:Take PicturesFor the sake of this tutorial, let's take this statue as a subject. You want to click about 50 images of the statue placed stationery on a stable surface and illuminated uniformly. Make sure you cover the subject from all angles and height as far as possible.Step 2:Upload the images to google driveThis one needs no explanation. I uploaded the images to a folder named Doll.Step 3:Goto the colab link below.The first link in the description is the link to google colab script, using which we convert these pictures to 3d model. Step 4:Setup the scriptSequentially run the code cell one by one.The first cell downloads and installs the meshroom which is a free open source photogrammetry software to your virtual machine on the cloud.The second cell will connect this machine to your google drive.Please connect the google drive to the one containing the pictures.The last cell will initiate the actual process needed for 3d reconstruction. This cell requires a setup. The setup demands you to assign the correct path to pictures of the subject uploaded to your drive as well as the output destination folder where the program will store the 3d model.Recall your images are in the Doll folder. Assign the path to this folder as input.Also for the output destination, I created a new folder named Output_folder. This should show up in colab. Finally mention the path to this folder as output.Step 5:Execute the final code cellsThis process will require you to wait. In my case it took just over 2 hours. Step 6:Download the output (3d model)Go to the output directory you set in the google drive and download this folder.Step 7: Finally, check the generated model.Depending on the output, you might need to clean the model. You could do it easily using blender which is a free open source, cross platform 3d software.
Covid-19 Recovered Cases | 3d graph visualization Cominlooks 2020-05-01 | The 2019–20 coronavirus pandemic is an ongoing pandemic of coronavirus disease 2019 (COVID-19) caused by severe acute respiratory syndrome coronavirus 2 (SARS‑CoV‑2).This 3d graph visualization presents the curve of various countries for conformed corona recovered cases.Hope you are healthy and enjoyed this video. Please like, share, subscribe to this channel showing your support for upcoming videos, because it does take effort in making these videos. Code/Script:https://github.com/AwaleSajil/Covid-1...Datasource:https://data.humdata.org/dataset/nove..."Art of Silence - by Uniq" is under a Creative CommoMusic:ns license (Creative Commons – International Recognition 4.0 – CC BY 4.0)
Corona virus Conformed Deaths | 3d Graph visualization Cominlooks 2020-04-26 | The 2019–20 coronavirus pandemic is an ongoing pandemic of coronavirus disease 2019 (COVID-19) caused by severe acute respiratory syndrome coronavirus 2 (SARS‑CoV‑2).This 3d graph visualization presents the curve of various countries for conformed corona deaths.Hope you are healthy and enjoyed this video. Please like, share, subscribe to this channel showing your support for upcoming videos, because it does take effort in making these videos. Code/Script:github.com/AwaleSajil/Covid-19-cases_scriptDatasource:data.humdata.org/dataset/novel-coronavirus-2019-ncov-cases"Art of Silence - by Uniq" is under a Creative CommoMusic:ns license (Creative Commons – International Recognition 4.0 – CC BY 4.0)Thank you
Covid-19 Infection trend | 3d Graph visualization | Script included in description Cominlooks 2020-04-25 | The 2019–20 coronavirus pandemic is an ongoing pandemic of coronavirus disease 2019 (COVID-19) caused by severe acute respiratory syndrome coronavirus 2 (SARS‑CoV‑2).This 3d graph visualization presents the curve of various countries for conformed corona cases.Hope you are healthy and enjoyed this video. Please like, share, subscribe to this channel showing your support for upcoming videos, because it does take effort in making these videos. Code/Script:github.com/AwaleSajil/Covid-19-cases_scriptDatasource:data.humdata.org/dataset/novel-coronavirus-2019-ncov-cases"Art of Silence - by Uniq" is under a Creative CommoMusic:ns license (Creative Commons – International Recognition 4.0 – CC BY 4.0)Thank you
Animation: Story of Corona Virus Cominlooks 2020-04-06 | Corona in spaceas it spreads and builds its basehaving enough time to grow facial hair and smoke cigar all the while bringing nations under its feetA smoke alarm goes off, fireman comes to the rescue burns the corona downRejoice the victory until it lasts It is just the tip of the icebergthe number showsfrom black and white to RBG, it thrivesBut gazing at the big picture,it was all inside a box called quarantineunable to reach out to the universeYou see there is a happy endingWe stayed put and we stayed healthyVideo is not intended to take away the seriousness of the issue.Music:Epic Drama Cinematic Trailer
