Huddly’s Post

Like everyone else for the last 48h I have been trying Flux.1 Pro, the new open-source GenAI model by Black Forest Labs, a team composed of some of the original engineers of Stable Diffusion. I am not going to focus on benchmarking whether it really is better than Midjourney V6.1 but I am going to touch point about « Prompt Adherence ». If you aren’t familiar with the term, it is the capacity of a GenAI system to understand, decompose, and recompose precise instructions in a prompt. When I started using Jax Diffusion now 2.5 years ago, I created a post-apocalyptic supernatural character called KIDD-0 through photobashing and 3D which I was then running through AI passes. I figured it would be interesting to test the prompt adherence of Flux by trying to recreate the very same character with just a prompt. If you are familiar with Midjourney or Stable Diffusion, you know Prompt Adherence got better over time but never has been good enough to understand complex instructions on clothing, hair, colors, etc. Yes you can train Loras for better consistency but controlling a composition is its own thing. Conclusions: Flux really is amazing at understanding complex instructions. A little photoshop, Firefly, Magnific and I managed to get the character I imagined years ago. Looking forward to seeing where this new player goes.

Multimodality - from pure, raw imagery to the evolving narrative around expanded ways to represent cognitive signal - is the single most exciting element of AI I’m looking forward to the most!

Introducing #SDXLTurbo: A #RealTime Text-to-Image Generation Model! Check it out in the #video below! Links to Stability AI's files, papers, code, and more, are in the comments. SDXL Turbo achieves state-of-the-art performance with a new distillation technology, enabling single-step image generation with unprecedented quality, reducing the required step count from 50 to just one. SDXL Turbo is based on a novel distillation technique called #AdversarialDiffusionDistillation (ADD; insert tech worker jokes here), which enables the model to synthesize image outputs in a single step and generate real-time text-to-image outputs while maintaining high sampling fidelity. SDXL Turbo iterates on the foundation of SDXL 1.0, and by incorporating ADD, SDXL Turbo gains many advantages shared with #GANs (#GenerativeAdversarialNetworks), such as single-step image outputs, while avoiding artifacts or blurriness often observed in other distillation methods One can download the model weights and code on #HuggingFace, currently being released under a non-commercial research license that permits personal, #noncommercial use. One can also test SDXL Turbo on Stability AI’s image editing platform #Clipdrop, with a beta demonstration of the real-time text-to-image generation capabilities via their browser.

Generative models have stunned us all with their capabilities, becoming incredibly popular in recent years. As we've navigated through this evolving landscape, we've learned that trial and error often lead us to excellence. But what if we could streamline this process? How many attempts should it take to perfect the art of image synthesis? That what we talk about in our report: 🔗 Read the full report at: https://lnkd.in/dGaJAtwz TL;DR More or less 5 trials are enough. #GenerativeAI #DiffusionModels #AIResearch Unrelated, very cool video on MAMBA here: https://lnkd.in/dzzQZA2s

If you're a data professional looking to the great resource to dive into the world of Multimodal and Large Multimodal Models (LMMs), this one's for you! https://lnkd.in/d3yrw8eZ Read this article which help you to discover the potential of working with diverse data modalities like text, images, videos, and audio. It's the secret sauce for AI in the real world across healthcare, robotics, e-commerce, retail, gaming, and more. And let's give a big shoutout to Chip Huyen for her enlightening research paper (article) that inspired me to share this post! Keep the knowledge flowing, and remember, your feedback is invaluable! #GenAI #LargeMultimodalModel #AIResearch #NLPEngineering #CareerGrowth #transferknowledge #transformlives

OpenAI just released atext-to-video model, V-JEPA, taking a step towards Yann LeCun's vision of advanced machine intelligence (AMI). It excels at detecting and understanding detailed interactions between objects, contributing to a more grounded understanding of the world for machines. Released under a Creative Commons NonCommercial license. Key bits: ✅ V-JEPA predicts missing or masked parts of a video in an abstract representation space, improving efficiency by 1.5x to 6x compared to generative approaches. ✅ It's pre-trained entirely with unlabeled data, requiring labels only for task adaptation. ✅ V-JEPA's masking methodology and self-supervised learning approach contribute to its efficiency and robustness. ✅ It introduces a novel frozen evaluation technique, enabling efficient adaptation to new tasks without full fine-tuning. Future directions include incorporating audio for a multimodal approach and extending its capabilities for longer time scales and planning tasks. As a research model, V-JEPA holds promise for various applications, from embodied AI to contextual assistants for AR glasses. Released under the CC BY-NC license, it invites collaboration and further exploration in the AI community. Read the paper: https://lnkd.in/g8MYpG5s Get the code: https://lnkd.in/gVGbRpTU #OpenAI #VJepa #TextToVideo #MachineIntelligence #AIResearch #AIInnovation #FutureTech #SelfSupervisedLearning #MultimodalAI #ResearchCollaboration https://lnkd.in/gZkG8rJA

“Our mission is to help you solve your problem in a way that is super cost-effective and available to as many people as possible.” Tenstorrent CEO Jim Keller on why #AI models are graphs and how we built our software and hardware to give you a fundamental advantage at this year's AI Hardware Summit. #artificialintelligence #machinelearning #compute #tenstorrent #hardware

Check out this recording of CEO of Tenstorrent Inc. , Jim Keller's keynote at last month's AI Hardware & Edge AI Summit. Definitely among the most engaging and energetic keynotes of the conference, not one to miss! AI Hardware & Systems, #aihw #machinelearning #chipdesign #semiconductor

⁉ Why use AI for Video Compression? 🙋♂️ Perhaps ... to encode 3x faster, get higher video quality (6.7 VMAF points more), at the same bitrate? Yes - that's right. Visionular's AI-powered Aurora5 encoder screams past its competition in a recent benchmark. 🔍 For the same bitrate (2mbps), → Popular HEVC Encoder: VMAF score = 88.24 at 98 FPS. → 𝐀𝐮𝐫𝐨𝐫𝐚5: 𝐕𝐌𝐀𝐅 𝐬𝐜𝐨𝐫𝐞 = 94.89 𝐚𝐭 𝐚 𝐬𝐭𝐮𝐧𝐧𝐢𝐧𝐠 301 𝐅𝐏𝐒. That's 3x faster, 6.7 VMAF points more, at the same bitrate 👊🚀 And, our #AI-based compression is consistent across Aurora5 (HEVC), Aurora4 (H.264/AVC) and Aurora1 (#AV1), offering unparalleled technical and financial advantages for your business. Ready to elevate your encoding game? Contact us (https://lnkd.in/dqWtcP8U) for a seamless transition via either AuroraCloud or our SDKs. Zoe Liu | David Lea | Bo Zhang | Will C. | Alex Steel | Thomas Davies | Lukasz Kubala | Dmitriy Sychov | Rahul Rathore | Krishna Rao Vijayanagar  | Ray Wang | Charles Yao | Meryl Gu #hevc #aitranscoding #aiencoding #av1 #avc #transcoding #encoding #videocompression

Introducing #SDXLTurbo: A #RealTime Text-to-Image Generation Model! Check it out in the video below!