Playing with diffusion models leads to the awareness of concepts without words to describe them. Today, words don't precisely control image generation, but I expect this to improve over time. Soon, we might be expanding our vocabulary by generating #aiart!
https://twitter.com/DiffusionPics/status/1620539111961497602
Our interaction changes with every new version of the AI product. How one prompts in #dalle2 is different from #midjourney. Each version of MidJourney requires different prompting. Each finetuned model of #stablediffusion also prompts differently. Prompting has no standards.
The constraints that drive the generation of an image depend on the relationship between many variables. The relationship of words in a prompt, the strength of knobs, the iteration between images, the iteration between models, and the overall sequence of itself.
Stuart Kauffman describes that going beyond the adjacent possible cannot be prestateable. His example is a screwdriver. Can we enumerate (i.e., prestate) all the uses of a screwdriver? A screwdriver may be designed for a specific purpose, we can't know what it can be used for.
Deep learning research groups that are inventing networks every day, do not have the capacity to know what the future impact of their discoveries. At best they provide some metrics comparing against standard use-loads.
Methods may appear to do the same thing, and it's usually the case that methods are compared against other methods. But in practice, different methods have strengths. Textual inversion, dreambooth, and LORA all appear to do the same thing. But they behave subtly differently.
The same can be said about #dalle, #midjourney and #stablediffusion. There is no universal best tool. They execute differently, and there is value in knowing how one specific tool is better in a specific context.
#midjourney and #stablediffusion are smaller networks compared to #dalle, but they are more capable in their distinct ways. This observation will also translate to language models. Smaller networks like @anthropicAI Claude may be better that @openai ChatGPT.
Every new technology leads to an extreme variety of alternative solutions. At the dawn of the PC revolution, Apple had many competitors (i.e., PET, TRS-80, Sinclair, Commodore, Atari, etc.) What differs today in AI is that you can mix and match. en.wikipedia.org/wiki/History_o…
Today's Deep Learning has a peculiar kind of modularity. This modularity wasn't as obvious a few years ago. This is because generative AI fundamentally offers greater modularity than classification AI. medium.com/intuitionmachi…
The predominant error in the evolution of deep learning is that for years the emphasis on good classifiers dominated the landscape. This is because classification is easier to measure and easier to create a leaderboard competition. medium.com/intuitionmachi…
Language, evolution, and technology are all generative processes. Dalle-like and GPT-like systems are generative processes. These processes are not siloed and thus allow unexpected new inventions. They are open-ended by their nature.
@threadreaderapp unroll
• • •
Missing some Tweet in this thread? You can try to
force a refresh
