SDXL's VAE can be modified to use neighbourhood attention, accelerated by NATTEN. in fact, you can remove attention entirely and images look pretty similar. left = global self-attention right = no attention! thanks Rivers Have Wings for the idea! github.com/Birch-san/sdxl…

What papers should I read to expand my knowledge of Transformers? Please send links in the comments and write why this paper is worth reading. Thanks for your help!

In 2008, Google Maps launched in India. But we quickly ran into a problem: Nobody used street names. And street names were the foundation of Google Maps. The team had to make some big adaptations. 15 years later, the changes have stood the test of time. Here's how the team

Readout Heads look like the next evolution of guided image generation! Like ControlNet, readout heads can be used for pose, depth, or edge-guided generations. But compared to ControlNet models, they're much more lightweight. Let's take a closer look:

We introduce W.A.L.T, a diffusion model for photorealistic video generation. Our model is a transformer trained on image and video generation in a shared latent space. 🧵👇

I have just published an article explaining Segment Anything Model on medium link.medium.com/V6Yift8UPFb . Hope y'all like it.

Looking for deep learning project ideas. Topics I am interested in includes Domain Translation and Explainability. Reference papers would be appreciated😅. I am a final year undergrad so do show some mercy 😂😅.

In 2022, I used to launch startups in the void. - Post daily on 𝕏 - Make a product demo video - Share the Product Hunt link for support But I got 0 sales...

If you are curious about how non-max suppression works, some time ago, I made an infographic explaining this algorithm. 7/n

Working at really great efficiency, wrote two articles last week😅😱 I-JEPA: medium.com/@utkarsh135/i-… and MAE enable efficient knowledge distillers: medium.com/@utkarsh135/ma…

🎨 Introducing Follow-Your-Click: Transforming the way we animate images with just a click and a short prompt! New ImageToVideo model. Say goodbye to moving entire scenes and hello to precision and creativity.

✨ For the last few months I have been reverse engineering Magnific AI's famous upscaler It uses MultiDiffusion, ControlNet tiles and details LoRas In true AI spirit, I am open sourcing it for everyone to use for free in your apps It's designed to easily control the amount of

Apples, oranges, and the future of 3d ;-) Realtime, instant feedback, real-time lighting, and effects in a more painterly, sculptural, intuitive workflow. An underlying game engine lets you make full games, interactive movies, or art installations. AI takes care of the final

FRESCO combines ControlNet with Ebsynth! The result: Improved spatial and temporal consistency for video-to-video 🔥 Links and 5 cool examples:

A very Good Saturday Morning🙏 Sharing a beautiful video received on WhatsApp👌💎💪Enjoy the Life in full swing🙂💪 Love Life, Live

Entrepreneurs face pressures most people won’t understand. 20 sentences to read when you’re feeling stuck: 1. You've survived 100% of your worst days so far. This too shall pass.

Gall's Law is one of those statements that just keeps hitting harder and more deeply the more time I spend with it. At first it seemed false, then a truism, then an interesting insight, and now a foundational belief about the world.

I rewatch this once a month for motivation