🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

[1/10] 🚨 We present our recent Snap Inc. project: Stable Flow --- A training-free method that performs various types of image editing operations (e.g., non-rigid editing, object addition and replacement) using flow models. Project page: omriavrahami.com/stable-flow

.

🎥 𝗩𝗕𝗲𝗻𝗰𝗵 𝗔𝗿𝗲𝗻𝗮: 𝗪𝗮𝘁𝗰𝗵 𝗔𝗜-𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗲𝗱 𝗩𝗶𝗱𝗲𝗼𝘀 𝗜𝗻𝘀𝘁𝗮𝗻𝘁𝗹𝘆 🎥 ✅ 180,000+ AI-generated videos, 40+ models (and growing) ✅ You can optionally vote for your preferred outputs Try it here: huggingface.co/spaces/Vchitec…

Our **Flow Matching Tutorial** from #NeurIPS2024 is now publicly available: neurips.cc/virtual/2024/t… Heli Ben-Hamu Ricky T. Q. Chen

Super excited to share some recent work that shows that pure, text-only LLMs, can see and hear without any training! Our approach, called "MILS", uses LLMs with off-the-shelf multimodal models, to caption images/videos/audio, improve image generation, style transfer, and more!

Introducing “Diffusion Autoencoders are Scalable Image Tokenizers” (DiTo). We show that with proper designs and scaling up, diffusion autoencoders (a single L2 loss) can outperform the GAN-LPIPS tokenizers (hybrid losses) used in current SOTA generative models. (1/4)

Tokenizers in image/video generation are way understudied! "Standard" recipe: combinatorial search over different losses using a plethora of models DiTo is our attempt to break-away from this: simpler, scalable, and theoretically sound! Idea: Use diffusion to learn the tokens

Flow and diffusion-based video models are typically trained to denoise pixels. With a similar FLOP budget, they can simultaneously denoise additional video derivatives, such as optical flow, which captures motion more explicitly. This can significantly enhance motion and physics

Yuval Kirstain Wow what a great idea! Obvious in hindsight, as the best often are.

Meta just dropped VideoJAM Joint Appearance-Motion Representations for Enhanced Motion Generation in Video Models comparison with openai sora and kling

This is extremely cool! They find diffusion loss is not very sensitive to motion. Thus they fine-tune videogen models with additional explicit motion prediction, making the model generate much more coherent videos. Also, Hila has been doing consistently good work, follow her!

Thrilled to share our latest research on fundamental variable multi-layer transparent image generation, inspired by Schema Theory! ✨ ART enables precise control and scalable layer generation—pioneering a new paradigm for interactive content creation. 🚀 art-msra.github.io

🔥 Llama-level performance with <0.1% of the training data 🔥 Together with Cartesia, we introduce Llamba—a family of recurrent language models distilled from Llama-3 into Mamba. ⚡ Sizes: 1B, 3B, 8B 🚀 Optimized for speed & on-device efficiency Details here 🧵👇

Qiying Yu and team just dropped the DAPO algorithm (decoupled clip and dynamic sampling policy optimization)! DAPO-Zero-32B, a fully open-source RL reasoning model, surpasses DeepSeek-R1-Zero-Qwen-32B, and scores 50 on AIME 2024 with 50% fewer steps. It is trained with

Test-Time Training (TTT) is now on Video! And not just a 5-second video. We can generate a full 1-min video! TTT module is an RNN module that provides an explicit and efficient memory mechanism. It models the hidden state of an RNN with a machine learning model, which is updated

The Transformer–SSM retrieval gap is driven by just a few heads! SSMs lag on tasks like MMLU (multiple-choice) and GSM8K (math) due to in-context retrieval challenges. But here’s the twist: just a handful of heads handle retrieval in both architectures. What we found 👇 1/

Padding in our non-AR sequence models? Yuck. 🙅 👉 Instead of unmasking, our new work *Edit Flows* perform iterative refinements via position-relative inserts and deletes, operations naturally suited for variable-length sequence generation. Easily better than using mask tokens.

Exciting news from #ICML2025 & #ICCV2025 🥳 - 🥇 VideoJAM accepted as *oral* at #ICML2025 (top 1%) - Two talks at #ICCV2025 ☝️interpretability in the generative era ✌️video customization - Organizing two #ICCV2025 workshops ☝️structural priors for vision ✌️long video gen 🧵👇

[1/n] New paper alert! 🚀 Excited to introduce 𝐓𝐫𝐚𝐧𝐬𝐢𝐭𝐢𝐨𝐧 𝐌𝐚𝐭𝐜𝐡𝐢𝐧𝐠 (𝐓𝐌)! We're replacing short-timestep kernels from Flow Matching/Diffusion with... a generative model🤯, achieving SOTA text-2-image generation! Uriel Singer Itai Gat Yaron Lipman