The attention layers in the VAEs for FLUX, Stable Diffusion 3.5, and SDXL don't do anything. You can ablate them with almost no effect. At first I thought they might be involved in some clever circuitry—maybe moving global information—but no they're just flailing around doing

There's a lot more to say, but I'll stop here for now. Check out the paper! openreview.net/forum?id=7lUdo…

it’s over turns out the rl victory lap was premature. new tsinghua paper quietly shows the fancy reward loops just squeeze the same tired reasoning paths the base model already knew. pass@1 goes up, sure, but the model’s world actually shrinks. feels like teaching a kid to ace

RampenSau by David Aerne meodai.github.io/rampensau/

We are pretty excited to announce the latest updates on our space on Hugging Face 🔥. Previously, subject consistency would sometimes break when generating the subject in various angles. Now do it with different backgrounds while keeping its consistency throughout angle changes

Gradient + Blur — Woww 💙 #figmaconfig2025

GPT-4o + JSON = next-level visuals with precision and style! Prompt 👇

The same landing page, now with animations 👀✨

Flow-GRPO Training Flow Matching Models via Online RL

AirBnB icons but make them Dieter Rams. Prompt below, I think you can load this in 4k too..

I've been quietly generating a collection of 3D icons for my WIP project Thiings over the last few weeks. It felt like a good time to open the collection up (*cough* airbnb *cough*) All icons available for download as PNGs with transparent backgrounds. Link below 👇

The problem with oAI: hiring product people has zero understanding of the problem. They try to maximizes vibe coding, targeting engineers who prefer full control and precision. Unless GenAI is 100% reliable, it isn’t suitable for professional tools. youtube.com/live/hhdpnbfH6…

run action interaction ⎯⟡°

so...

NEW: Visit our Self-Serve Portal and get commercial licenses for our open weights models with only a few clicks. One portal. Transparent pricing. Retrieve a commercial license in minutes. Read more about FLUX.1 Kontext [Dev] in our deep dive bfl.ai/announcements/…

Had the honor to present diffusion transformers at CS25, Stanford. The place is truly magical. Slides: bit.ly/dit-cs25 Recording: youtu.be/vXtapCFctTI?si… Thanks to Steven Feng for making it happen!

During training, diffusion models are being taught to be effective denoisers, like Associative Memory systems. At what point do these models stop being denoisers and behaving like data generators? To learn about how these models arise from being Associative Memory systems to

📢 The Gemini Embedding text model (gemini-embedding-001) is now generally available in the Gemini API via Google AI Studio. It supports 100+ languages and uses Matryoshka Representation Learning for flexible output dimensions, allowing devs to scale down from 3072 dimensions.

If you only want dominant colors - just do a quick k-means on the 25 pixels, add some contrast, and get a pretty good color palette 🎨 You can run the same algorithm on groups of images and use the output to influence your shaders 🧑‍🎨 (5/6)