The themes chosen by Re-impressionists were also in challenge to the #AI-based censorship, exploring so-called NSFW themes that would be aggressively censored by cloud services which monopolized terra- and peta-byte generative models. "Not Safe For Who?"
@alexjc
-
Adversarial Pixel Patterns Outside Natural Image Manifold
By
–
In particular, the pixel patterns used were specifically not on the "manifold" of natural images — thus making it difficult for algorithms such as Latent Diffusion to generate good results. Ironically, optimization techniques were used to find these adversarial pixel patterns!
-
Pixel patterns cause image blur for corporations using uncurated datasets
By
–
The use of these pixel patterns resulted in increasingly blurred images throughout the 2020s for corporations still using uncurated and infringing web-scale datasets — as opposed to the industry standard datasets based on opt-in consent.
-
AI Art Prompts Artists to Create New Distinctive Styles
By
–
Like the Impressionism movement that emerged with the popularisation of photography (increasingly used to "capture" realistic scenes), artists in the 2020s felt the need to define new styles to distinguish themselves from the art produced by #AI generators.
-
Re-Impressionism: Digital Art Movement Against AI-Generated Art
By
–
Re-Impressionism is a digital art movement that became popularised in 2023, as a response for — and in opposition to — the trend of #AI-powered art. Its core technique is a modern version of pointillism, where intricate pixel patterns are used to depict scenes.
-
Consent and Fair Compensation for Artists in AI Training
By
–
This is a good thread! Almost all replies about consent being opt-in, which is the right thing to do both from the artists perspective and from the companies they are licensing their dataset too. Users will win that one at this rate… Then it's about getting them paid!
-
Lack of Transparency and Editing Challenges in AI Models
By
–
Not necessarily overfitting, but lack of transparency and easy editing are the major things. See previous thread:
-
Hybrid AI Systems More Practical Than Full LLM Retraining
By
–
I love the dramatic futurism here! But it's more likely they will be hybrid systems. You don't want to retrain a LLM everytime a Wikipedia page is changed, or a news item is published. (Plus, Google would be in the best place to deliver this new system, incrementally.)
-
T5 and CLIP Embeddings: Superior Combined Performance Analysis
By
–
They ablate the embeddings in the paper and found that T5 and CLIP together are superior to either on their own. Not clear just how much difference it makes on text though… Imagen is bigger than SD, hard to disentangle the size from embedding on text. Good hypothesis though!