Image created with OpenAI GPT-Image-1. Image prompt: TikTok LIVE phone-screen POV, floating hearts & spinning album art, Retro8bit pixel-art filter with CRT scanlines, featuring split-screen showing image, text and audio wave overlay; soft-glow studio lighting, photoreal 8k
Ring Video Descriptions deliver real-time, Gen AI descriptions of what’s happening https://www.aboutamazon.com/news/devices/ring-video-descriptions-gen-ai
NO MORE SEO VideoPrism by @GoogleDeepMind is 🔥 it’s a versatile video encoder that can be plugged into text encoder or LLMs the authors first train a CLIP-like video-text model, then distill video encoder in masked manner to VideoPrism 😮 all models with A2.0 license on @huggingface 🤗 https://x.com/mervenoyann/status/1937572802896200181
https://research.google/blog/videoprism-a-foundational-visual-encoder-for-video-understanding/
Gemini is good at processing video (using frequent screenshots & audio transcripts). I gave Gemini a video on a historical recipe, it was able to find visual elements not mentioned in the transcript. It is not hallucination-free, but there are lots of new use cases for screening https://x.com/emollick/status/1936446885029372106
Gemma 3n is out, with day-0 MLX support 👏 https://x.com/awnihannun/status/1938283694416077116
Introducing Gemma 3n: The developer guide – Google Developers Blog https://developers.googleblog.com/en/introducing-gemma-3n-developer-guide/
We’re fully releasing Gemma 3n, which brings powerful multimodal AI capabilities to edge devices. 🛠️ Here’s a snapshot of its innovations 🧵 https://x.com/GoogleDeepMind/status/1938278533517746686
We’ve taken community feedback very seriously, and that’s why for Gemma 3n launch we’re so proud to partner with so many in this amazing ecosystem Thanks to @huggingface, @ollama, @Prince_Canuma for MLX, @UnslothAI, @ggerganov llama.cpp/GGUFs, @NVIDIAAIDev, @kaggle,”” / X https://x.com/osanseviero/status/1938349897503412553
11/ TL;DR MJ has dropped a decent model with unlimited video gen for $60/month. if you’re all about the MJ aesthetics, doing abstract generations (sans text), and don’t mind using MMAudio or similar to add audio in post, this is a great addition to the toolkit. https://x.com/bilawalsidhu/status/1935528281031462945
MidJourney Video 1/ First off. It’s fun to just click through your MJ catalog and see it come to life. – No text to video; only image to video for now – Works with MJ or uploaded imagery – Can choose high/low motion + auto or custom prompt – Can extend clips 4x – SD output at 24 fps; no upscaling https://x.com/bilawalsidhu/status/1935527429768163481
MidJourney Video 4/ Pretty good with motion graphics and AR visuals. Lack of text rendition (a weakness for MJ in general) comes through. Would not recommend for titles. But you can still get some beautiful abstract visuals (e.g head locked AR gen on the left, multi-monitor generations on right) https://x.com/bilawalsidhu/status/1935527747725709668
MidJourney Video 6/ Muzzle flashes look pretty good. But I had a very hard time getting shell casings to work properly. Let me know if you find a good prompt to achieve this, because I couldn’t. https://x.com/bilawalsidhu/status/1935527929569755261
MidJourney Video 7/ MJ video nails that high end unreal engine “”rendered”” look. Fisheye lens distortion and sweeping camera move is nice. But notice how wonky all the cars in the scene look. As it stands, I don’t think we’ll be pulling any 3d objects or scenes out of this video model. https://x.com/bilawalsidhu/status/1935527993830686966
MidJourney Video 8/ Some generations get this weird “”unsharp mask”” look (a technique for sharpening) as the generation progresses. Lmk if you spot it too. https://x.com/bilawalsidhu/status/1935528060612395421
MidJourney Video 10/ MJ video does seem like an amazing tool to make abstract visual elements you composite elsewhere. I hope they remove the extend duration limit (20s / 4 times max) because it could be an amazing tool for screensavers, music videos and concert visuals. https://x.com/bilawalsidhu/status/1935528210588213670
MidJourney Video 2/ Fast generation time. Works well for that wide angle vlogging style. You can extend any clip 4x. Two great examples below. Of course, it’s begging for dialogue. Sure, you can add the facial performance in post – but it won’t look half as good. Veo 3 has spoiled us here. https://x.com/bilawalsidhu/status/1935527555404271877
MidJourney Video 3/ MJ video does okay in my handshake test (homie on the left really went in hard lmao) Physics is a weakness of this model — doesn’t matter if it’s soft-body or rigid-body subject matter. Might get slightly better as user ratings roll in, but still far behind the SOTA. https://x.com/bilawalsidhu/status/1935527672484179979
MidJourney Video 5/ The dinosaur test comes next. Movement looks decent, but the rest of the physics in the scene are all over the place. The slipping tanks in the background reminds me a bit of Sora. Relative scale and relative motion is pretty wonky. https://x.com/bilawalsidhu/status/1935527810564767970
MidJourney Video 9/ Testing fluid simulations here. Not only is it pretty far from SOTA, sometimes I get generations with this stop motion-like choppy FPS look (e.g. wine glass on right). https://x.com/bilawalsidhu/status/1935528134113407408
The new Hailou 02 AI video model really does seem to have made huge strides in the “”gymnastics problem”” where fast flipping motions lead to distortion Here are the first three results of the “”a man in elaborate robes does a backflip while holding two pool noodles”” (a hard test!) https://x.com/emollick/status/1936091679850705019
25MM views on @TEDTalks instagram 🤣 … i think the most ever? so crazy! 🙃 … this tech is comin’ folks. hello from production heaven in taipei! 👋😊 https://x.com/jasonRugolo/status/1875047066256515229
カプセルインタフェース：視聴覚や動き，力加減を伝え，全身リアル体験を実現 – CapsuleInterface: Full-Body Experience via Senses and Motion – YouTube https://www.youtube.com/watch?v=8a46Uap367k
Today we’re introducing you to the future of video. The world’s first Creative Operating System, we call it the HeyGen Video Agent. Upload a doc, some footage, or even just a sentence. It analyzes your input. Finds the story. Writes the script with taste. Selects the shots https://x.com/joshua_xu_/status/1938252187941122091
Reddit considers iris-scanning Orb developed by a Sam Altman startup | Semafor https://www.semafor.com/article/06/20/2025/reddit-considers-iris-scanning-orb-developed-by-a-sam-altman-startup
Day 2/5 of #MiniMaxWeek: Introducing Hailuo 02, World-Class Quality, Record-Breaking Cost Efficiency 🎥 – Best-in-class instruction following – Handles extreme physics (yes, it does acrobatics 🤹) – Native 1080p https://x.com/MiniMax__AI/status/1935026724468871550
Amidst the IO/IYO controversy, here’s Jason Rugolo’s demo from TED last year. An “AI audio computer” that edits soundscapes in real-time — solving the age old cocktail party problem. This video got a whopping 25M views and 770k likes on IG. https://x.com/bilawalsidhu/status/1937327052845150554
Phew, it’s been a while. timm 1.0.16 released today to provide the image encoder for Gemma 3n. Additions kept on stacking so I haven’t had chance to finalize a release since the last one to support SigLIP-2 backbones. Lots of stuff in there: * Gemma 3n encoder (via a”” / X https://x.com/wightmanr/status/1938311403934519807
RT @GoogleDeepMind: We’re fully releasing Gemma 3n, which brings powerful multimodal AI capabilities to edge devices. 🛠️ Here’s a snapshot…”” / X https://x.com/slashML/status/1938394979727999455
RT @osanseviero: I’m so excited to announce Gemma 3n is here! 🎉 🔊Multimodal (text/audio/image/video) understanding 🤯Runs with as little as…”” / X https://x.com/algo_diver/status/1938374626910060782
RT @reach_vb: Google COOKED yet again – Multimodal Gemma3n 4B and 2B now available in Transformers, vLLM, MLX AND Llama.cpp 🤯 The model ca…”” / X https://x.com/reach_vb/status/1938476208330866751
RT @simonw: I’m really impressed by the new Gemma 3n I tried a 7.5GB model from Ollama and a 15GB model through mlx-vlm – they seem very c…”” / X https://x.com/osanseviero/status/1938581225452486911
RT @UnslothAI: Run Gemma 3n locally with our Dynamic GGUFs!✨ @Google’s Gemma 3n supports audio, vision, video & text and the 4B model fits…”” / X https://x.com/osanseviero/status/1938307534840074522
KerasHub lets you use any Hugging Face checkpoint for all top models like Llama, Gemma, Mistral, etc… Run your workflows in JAX, PyTorch, TensorFlow – inference, LoRA fine-tuning, large-scale training from scratch Blog post: https://x.com/fchollet/status/1938208330062655678
Reid Hoffman Invests in Ultrasound and AI-powered Brain Startup Sanmai – Bloomberg https://www.bloomberg.com/news/articles/2025-06-23/reid-hoffman-invests-in-ultrasound-and-ai-powered-brain-startup-sanmai?embedded-checkout=true
Pretty Insane – SoTA Text to Speech model capable of English AND Hindi – 3B Llama backbone – Apache 2.0 licensed 🔥 > Sub 80 ms latency > Supports both English, Hindi including code-mix > Runs in a free google colab too 🤯 Best part: They’re actively working on other languages https://x.com/reach_vb/status/1937903143565885878
Try on looks and discover your style with Doppl https://blog.google/technology/google-labs/doppl/
Retrieving text chunks is the bread and butter of RAG. We’ve released a recent feature in LlamaCloud that allows you to not only retrieve text, but image elements from documents as well. 1️⃣ You can index, embed, and retrieve embedded figures (charts, pictures) within a PDF and https://x.com/jerryjliu0/status/1936451556293104067
Most of the value in AI video won’t be captured by the creation tools. It’ll accrue to the platforms — X, YT, IG, etc. — where that content is distributed, ranked, and monetized. Can’t imagine it playing out any other way.”” / X https://x.com/bilawalsidhu/status/1935854180310434255
General-purpose multimodal reasoning systems struggle to flexibly handle diverse inputs and complex tasks with a unified framework. MEXA (Multimodal Expert Aggregator), a training-free framework, solves this by dynamically selecting and aggregating specialized expert models https://x.com/rohanpaul_ai/status/1937451383218930044
Vision-Language Models (VLMs) struggle with genuine deep comprehension of long-form video content because existing benchmarks often test shallow details or rely on low-quality, automatically generated questions. This paper introduces Movie Facts and Fibs (MF^2), a new benchmark https://x.com/rohanpaul_ai/status/1937466483275366761
Excited to share the release of VideoPrism! 🎥 📏Generate video embeddings 👀Useful for classifiers, video retrieval, and localization 🔧Adaptable for your tasks Model: https://x.com/osanseviero/status/1937560015348597124
China unveils a new brain for humanoids. The Beijing Academy of Artificial Intelligence (BAAI) has announced RoboBrain 2.0, an open-source general-purpose AI model designed to power humanoids and other general-purpose robots. BAAI is already collaborating with 20 Chinese https://x.com/TheHumanoidHub/status/1934326215382569132
Helix neural network https://x.com/adcock_brett/status/1937577814875881889
Existing knowledge distillation for Vision-Language Models (VLMs) struggles when teacher and student models have different token types, limiting knowledge transfer. This paper introduces Generation after Recalibration (GenRecal), a novel general-purpose framework using a https://x.com/rohanpaul_ai/status/1936665454862360719
LeVERB is a VLA framework for humanoid whole-body control, combining a vision-language model and a low-level controller via a shared latent action space, trained entirely in sim, deployed zero shot. https://x.com/TheHumanoidHub/status/1937500208423923717
