Twitter isn't showing article titles anymore, so I made this tool that adds the title to the preview for you! Check it out ➡️ http://headline-image.vercel.app

R to @marissamary: From GPT4 is clearly generic. Realistic Painting is about shadow definition and being very precise shadow edge, shape and value, not about “darkening”. Narrow value realistic painting can be achieved.

R to @v1rtualmed1anet: 👆 Which one is better? ChatGPT’s answer consisted of generic tips that could have applied to any painting. Bard seemed to notice some particularities about the image and gave more specific advice.

R to @v1rtualmed1anet: 4. Tips for improving a painting Prompt by @marissamary: "How can I improve this painting to make it more realistic? Be specific."

R to @marissamary: AI art teacher: while true do Say( "Needs more contrast.") loop

R to @AmandaHilden: Haha thanks!

R to @marissamary: Shaders are hard, so it's nice the model was able to get close on simple patterns. Definitely going to test this out again in the future as the model improves.

R to @marissamary: I tried a few iterations for this more complex pattern. I told it to describe the pattern first, and it went a little off the rails by comparing it to a face. Left: input. Right: one of the results.

Trying to see if GPT-4V can generate GLSL fragment shader code based off of an image. Left: input. Right: result. It got the right idea here by using the distance function, but over-complicated it by adding a smoothstep blur. Examples from @bookofshaders

R to @marissamary: why is the left side of the cow in shadow? its hard to tell where the light source is coming from

R to @romanzubenko: yeah I'm really excited for the openai API to start supporting images- then you could have a realtime art or math tutor

R to @marissamary: AI as a tutor is my favorite use case for LLMs, now with multiple modalities this is getting insane

ChatGPT gave me the same critiques as my art instructor did on this study. Both said to smooth out the midtones and add another value between shadow and light. Including a persona in the prompt ("you are an art professor") seems to give much better results!

This is a very clever idea from @marissamary. Imagine asking ChatGPT to critique your photography as well... #dalle3 #ai #chatgpt

R to @mattcampux: Sure!

R to @marissamary: would you like to compare it to what a human might offer for painting advice, I can offer up some notes for comparison

I've been really excited about the potential for AI to make people better at painting, and I think we just made a big leap with GPT-4V. It identified the main thing to fix in the flower painting (darkening the shadows) and made multiple good suggestions for the cow painting 🤯

R to @marissamary: Overall, I think it's really promising that chatgpt can generate a starting UI from an image, and then iterate on it. Future things to try: - Getting chatgpt to write a detailed breakdown of an image in text, and then generate off of those rules - Getting more colour accuracy

R to @marissamary: Tested the colour detection and it's not quite right, so may have been a coincidence.