- AI For Humans - The Newsletter
- Posts
- 👗 🪨 Dwayne "School Girl of Rock" Johnson | Daily Drip
👗 🪨 Dwayne "School Girl of Rock" Johnson | Daily Drip
🤖 New Mind-Blowing Image-Gen (Multi-LoRA)
Kevin Pereira
07 Mar
🙌 Researchers have developed a technique called “Multi-LoRA Composition”, which rolls off the tongue and takes customization and detail to a whole new level. By integrating multiple Low-Rank Adaptations (LoRAs) in creative ways, they can generate insanely accurate and high-quality images that put traditional methods to shame.
Their examples, not mine. But, I’m not complaining. Look at how the model accurately stretches and contorts the source clothing over Dwayne’s body, even in a fully synthesized pose!
📸 To break it down, LoRAs are basically small models. Think targeted, finely tuned/trained datasets, to give you a consistent image style or character appearance…
🫠The secret sauce? Instead of just mashing LoRAs together and hoping for the best, they've got two “decoding methods” (LoRA Switch and LoRA Composite), which dynamically blend elements while keeping all the juicy details intact.
They're using GPT-4V to evaluate the compositions and image quality; this beast of an evaluator is showing a scary-good correlation with human judgment. đź‘€
🏀 The results? Their methods are straight-up dunking on the competition, especially when it comes to complex compositions.
Impressive. Sampling the character and clothing in one step, then applying an original background in the second. Check out the hand resting on the bookshelf: context awareness!!
💥 Just so we’re clear, I did not not write that final line about “straight-up dunking”! Claude 3 suggested I might, however, after reviewing some of my older blog posts. I am filled with a sadness which is, frankly, difficult to communicate. I regret, so many things…
If you dig these daily drips, or there’s something you think we should feature/discss, let us know, and please consider sharing this with your friends and frienemies:
Much love. Until tomorrow!
.k+g