Woah… This is cool. Today, I found some old photos that had gone blurry and faded. It was hard to see the memories clearly. I tried upscaling them with a few tools, but most either added watermarks or weren’t free. Then I came across LucidFlux - released today! This AI tool

🚀 SANA-Video: Linear Attention + Constant-Memory KV Cache = Fast Long Videos 💥 Key Features 🌟 🧠 Linear DiT everywhere → O(N) complexity on video-scale tokens 🧰 Constant-memory Block KV cache → store cumulative states only (no growing KV) 🔄 🎯 Temporal Mix-FFN + 3D RoPE

LucidFlux: caption‑free universal image restoration on Flux.1. Scaled real‑world HQ data- robust, high‑fidelity fixes with minimal overhead. Not so bad w2genai-lab.github.io/LucidFlux/

This is a very solid and promising research that scales consistency models to 10B+ video diffusion models. The combination of sCM and Variational Score Distillation is a very promising direction for few-step generation!

Ditto Scaling Instruction-Based Video Editing with a High-Quality Synthetic Dataset

today i had a talk in hkust gz, one friend asked me how come we can make the bet on scaling linear attention. my answer is more about the culture that i have been trying to make. admittedly it is too hard to change the mechanism which always rewards visible contribution and

Hey, our UltraFlux actually fits your definition of a native 4K text-to-image model. lol w2genai-lab.github.io/UltraFlux/

UltraFlux: Native 4K Text-to-Image Generation across Diverse Aspect Ratios This new Diffusion Transformer tackles coupled failure modes in positional encoding, VAE, & optimization with a data-model co-design. Expect sharper detail & stronger text-image consistency!

I am very excited to share that our paper, "One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Models" will be presented at #NeurIPS2025! Viacheslav Surkov is presenting it at #MexIPS2025: 📍𝐈𝐟 𝐲𝐨𝐮 𝐚𝐫𝐞 𝐚𝐭𝐭𝐞𝐧𝐝𝐢𝐧𝐠 𝐍𝐞𝐮𝐫𝐈𝐏𝐒 𝐢𝐧 𝐌𝐞𝐱𝐢𝐜𝐨

#ZImageTurbo のVAEをこれ、UltraFlux-v1に入れ替えるだけです。シャープさが全然違います！ < 拡大比較1枚目 huggingface.co/Owen777/UltraF…

久しぶりにVAEネタ・・・Z-Image TurboのVAEをUltraFlux-v1に変えると、確かにボケ感が軽減され、明らかに差がわかります😊

Boost Z-image quality with ease. UltraFlux fine-tuned VAE, trained on a 4K dataset. High speed, no additional costs. Sharpness goes brrrrr. huggingface.co/Owen777/UltraF…

From this project, I mainly learned three things: 1) Representation learning can fully emerge from generative objectives. In the language domain, this has almost become a consensus. However, in vision, discriminative representation learning methods such as CLIP and DINO still