• Alone or with a partner, come up with a creative application for using Stable Diffusion with p5 based on this starter code.
  • It is possible to combine this code with the OpenAI code we used earlier. For this you would have to add the openai.js file to the sketch and include it in index.html similarly to how localsd.js is included.
  • If the server stops working, please let @gohai know on Slack.

<aside> <img src="/icons/chat_lightgray.svg" alt="/icons/chat_lightgray.svg" width="40px" /> click here This is an AI helping you matche outfits for different occasions based on what you currently have. I had such kind of idea because I used to be troubled by not knowing what to wear everyday. To get the suggestions about the clothes pairing, I first send my request to ChatGPT 3.5 to know what could possibly be the other clothes matching what I have. Then I send the suggestions to the SD to generate the whole picture.

</aside>

• Images generated with Stable Diffusion are said to share a similar type of aesthetic and similar properties/flaws… Imagine yourself as an art critic who has to give a presentation about this image-making technique: Use Stable Diffusion to create 4 images that critique image-making with Stable Diffusion - for example, by exaggerating some of its unique quirks. Add a sentence or to two each of the images.

  

prompt: girl sitting under sunshine, warm, high quality, masterpiece, full detail, pastoral

negative prompt: chaotic, unmatched colors, catastrophic

critic: Sometimes we might need to avoid using it to generate objects that have logic (like a person have two eyes, one nose, one mouth, two arms, two legs and they should be on the right position) Or the picture we get could be very weird or even scary. 😰

prompt: an imaginable creature, 2D, flat color, comic, cartoon, interesting and funny, with very imaginable colors, simple background to show the full shape of the creature, high quality, large amount of details

negative prompt: chaotic, catastrophic, repetition, similar

critic: I used to think SD can only generate 3D images so this time I asked for 2D and flat color. It shows that we have different understanding towards imaginable color. For me, it should be bright or color with much variation, but the colors SD chose are deeper and less saturated. Trying to be mean, I also found the shadow is not accurate even though the shape didn’t need to follow any logic because of “imaginable”. Also, the details is not rich enough.

• Read Hito Steyerl’s Mean Images (New Left Review, Mar/June 2023, online). Rephrase a part of her critique of generative AI images with your own words.

<aside> <img src="/icons/chat_lightgray.svg" alt="/icons/chat_lightgray.svg" width="40px" />

1. Lack of Realism: It suggests that generative AI images are not realistic representations of actual objects but are statistical averages. They do not capture the richness and nuances of the real world.

2. The question above raises even worse issues – more training data could be focusing on the object human preferred. It's already impossible to make a resemblance, and with bias in the training data, it's even less so. How ridiculous is this world that mankind thinks it's centered on?

3. Exploitation of Labor: To prevent inappropriate images showing to the public, microworkers should identify and tag violent, biased and illegal material within datasets.

   Weirdly generated things + large amount of data= Physically and psychologically      damaged'
   

4. Lack of Ethical Consideration: We have no idea where those datasets come from. We don’t know the real purpose of them as the people controlling the AI keep it as a secrete. </aside>

• Come up with two questions about anything related to week 8’s content.

<aside> <img src="/icons/chat_lightgray.svg" alt="/icons/chat_lightgray.svg" width="40px" /> I’m still wondering how can we choose different lora and put them into p5js.

</aside>