#SyntheticMedia

[aniimaillussiya]

Individual prints are available at Pictorem: https://aniimaillussiya.pictorem.com

[swiftnliftnewsandarticle]

How do Deep Fakes work?

Deep fakes are synthetic media where a person's likeness is replaced in images or videos using deep learning. Here's a concise overview of how they work:

Data Collection: Gather images/videos of the target and source individuals from various angles and conditions.

2. Training the Model:

Neural Networks: Use deep learning models like Generative Adversarial Networks (GANs) or autoencoders.

GANs: A generator creates fake images, and a discriminator evaluates them, improving realism over time.

Autoencoders: Compress and then reconstruct the image from the source to resemble the target.

3. Face Swapping: Align and blend the source face into the target video, matching skin tone, lighting, and facial features.

4. Post-Processing: Enhance quality with smoothing and color correction. Synchronize facial movements and expressions to match the original video or audio.

5. Quality Assurance: Iterate training and refinement to improve realism. Human review helps spot inconsistencies.

6. Ethical and Security Concerns:

Deep fakes can spread misinformation and violate privacy.

Detection methods are being developed to identify deep fakes.

7. Tools and Technologies:

Software: Deep Face Lab, Faceswap, Adobe VoCo.

Libraries: Tensor Flow, PyTorch, Keras.

Understanding deep fakes is crucial for leveraging their benefits and mitigating their risks.

For More Information>>

[ldw1472]

◈🎧 | 💃오빠 나좀 바라봐!딥페이크 동영상🎵 🎶 ◈

[becominggiant]

SoDope by @technocalypse_now . . . #aiart #art #ai #digitalart #generativeart #artificialintelligence #machinelearning #aiartcommunity #abstractart #nft #aiartists #neuralart #vqgan #ganart #contemporaryart #midjourneyart #artist #nftart #artoftheday #newmediaart #syntheticmedia #aiartist #modernart #neuralnetworks #neuralnetworkart #abstract #styletransfer #stylegan #digitalartist #artbreeder https://www.instagram.com/p/Ci9RRTLOzNK/?igshid=NGJjMDIxMWI=

[jiparkitp]

Ji Text Generator - Final Project

undefined

What is it?

I made a clone version of me! You type starter words/sentence and it writes rest of the sentence based on the initial input.

Process

My goals for this project included:

Make a replica of my voice (text pattern).

Learn how GPT-2 model and Runway text training works

Collecting dataset

First, I started by collecting my text data. Mac makes it really easy to access your entire iMessage history. (I followed this very useful tutorial ) I was able to get about 20MB texts, going way back to my first archived texts from 2016.

Training

I trained on OpenAI's GPT-2 model with my imessage dataset. Everything was done via RunwayML.

Interface

I took a lot of reference from one of our class example by Yining. Im using pure javascript instead of p5js. You can view the full code here. The interface design is simple and clear. When user types in a sentence, it returns response. 

Takeaway & Next steps

I gained a lot of practice and knowledge from this process. It was definitely NOT fun to find out what the hell I was thinking from 2015 texts but I think the overall sentence makes some sense and it does really sound like me! I was shocked that it liked all the things I liked such as Korean food, kimchi, tea, fruit..etc. They are very accurate. Also, when I typed in Nicole, it returned ”Nicole’s house in Greenpoint is now a stone throw away from Williamsburg station”. This response is truly mind boggling for two reasons. First, it learned that Nicole lives in Greenpoint. And, it somehow learned the geography of Brooklyn. There isn’t a station called Williamsburg station but its a neighborhood right next to each other!

Pivots not failure!

Some responses were wonky and I think its because it was only trained with my texts. If there is a way to train with my voice(texts) but factor in the whole contexts such as entire conversation or my usual response to certain texts, I think it will generate more robust response. 

Things I want to improve...

Try fun little version of Turing test with my friends. See if my friends can find out if it is really me. 

Publish as a web site? 

Try with RNN and LSTM network model in ml5.js.

It would be interesting to see how my thoughts/voice changed throughout different stages of life.

I fell into some sort of existential crisis. I felt belittled (almost?) to see how predictable I was. Am I such a boring/predictable human being that it was able to just condensed me into a pretty much spot on algorithm??... But I had a lot of fun creating this project and I wish to continue! It will be my ongoing time capsule project :) 

[jasunfilter-blog]

Jasun Envirocare Plc is renowned as UK’s leading providers of quality air filters, as well as one of Europe’s leading manufacturer, supplier and online seller of air filtration products and hygienic water system treatment technology. Ever since it stepped into the air filter mainstream in 1972, Jfilters has largely contributed innovative concepts to improve the quality of air filtration and ventilation products, as well as state of the art technological advancements to purify commercial water methods globally.

[jiparkitp]

Final Project Proposal

Talk to me chat bot

Description Generate an online version of ‘me’ — Chatbot trained from my writings.

Inspiration Last week during class, I trained with data set of my texts and it generated pretty much accurate version of me. I want to explore if I can generate somewhat a clone version of me.

Visual reference It will be very simple chat like website

Audience I actually want to text it on my friends first without telling them and see if they notice. Kinda like Turing test per-se? I want users to have a smooth, human-like chat. I also want them to know about me through the chat as well. 

Challenge

I have quite a collection of my writings and texts already. Will have to work more to gather as much as possible

Programming - is it possible? I have used markov chain and done few chat bot following tutorial

[jiparkitp]

Week 2

Last week class, I was really inspired by gust speaker and fascinating examples of artworks/projects generated by ml5. In class, we tried generating images with ml5 plugin. Unfortunately, I found the photoshop plugin quite un-intuitive and hard to use. My adobe photoshop kept crashing while it was processing and when it was successfully executed, it took about 9 min for a small image.

For this week assignment, we were asked to generate images/video/texts using any ml5 models. During last class, we briefly talked about an artwork created by AI sold for £40,000 at Sotheby’s and decided to experiment with image generating models on ml5.

Big GAN

Find the p5js sketch here. (You need run ml5 simultaneously) BigGAN was interesting because it gets text input and responds with funky images like above. It was easy to use, fast and intuitive with p5js web editor.

SPADE - COCO This model translates user’s doodle into pre trained images. For example, if i draw a blob of wine bottle shapes or any shape with label “wine bottle”, it generates image accordingly.

[jiparkitp]

Intro to Synthetic Media Week1

First week with ml5 and Runway.

ml5

I wrote my first ml5 project! It is a simple image classifier. Im using video to capture images. —p5js web editor kept having ‘cross-origin’ problem with images. However, both video/image capture worked well when I ran it on a local server. I tried with couple objects and most of them were pretty correct. For example my dog’s photo was identified as ‘tissue paper’ (which I can’t really blame) and then it came back with my dog’s exact breed Maltese. 

Easy to use and understand. Documentation and references were clear and accessible. I’m excited to try new things. 

Runway

I used Keras-OCR Model that detects texts from images. I tried identifying the Houston-st station image and it came out as HOUSTON A ST. It identified small triangle dash between Houston and st as an ‘A’. This small hiccup was exciting because I was impressed by its ability to translate an abstract shape into a letter.

Runway was easy to use and understand as well. Surprisingly took less time than ml5. Could just be those few models I tried. 

[jiparkitp]

Week 3

This week’s assignment was to train a StyleGAN model with your own data using Runway.

Inspiration

I’ve been obsessed with Dog or Bread? meme for couple years now. It’s a famous memes that became an internet sensation, juxtaposing pictures of food with animals that look like that food.

I’ve discovered Cupcake on one of the pre trained model and decided to use my dog’s images as a data set to train.

Process

I had about 300 images of my dog in one folder. Before using the entire folder as data set, I manually deleted images that are messy (ex.if my dog is not the main subject, messy background..etc.) and saved images that are mostly just a closeup image of my dog’s face,

Results

I wasn’t super happy with the results but there were some fun images in the process. I think the training would’ve been more effective if the pre trained cupcake models were closeup of cupcakes. 

I tried training with portraits and here are the results!

[jasunfilter-blog]

Jasun Envirocare Plc retains its popularity as one of UK’s leading and largest air filter manufacturing companies, since its advent into the highly demanding and competitive industry. It is a renowned certified supplier and online seller of air filters and its various accessories for UK and France, as well as other European countries. One of its chief specialties is creating environmentally-friendly and effective techniques to improve standard air filtration and water hygiene systems globally

[jasunfilter-blog]

UK’s largest air filter manufacturer, supplier and online seller, Jfilters retains immense popularity with its past patrons and future customers. Jasun Envirocare’s goal has and always will be to introduce innovative concepts plus state of the art treatment that will improve not only the quality of air filtration and ventilation applications but also commercial water systems worldwide. We sell a widespread assortment of carefully monitored air filters on our online shopping website,