Hugging Face

Initializing each model in a DiffusionPipeline & then quantizing them to perform full inference is a painful dev experience 🥲 Today, we're shipping something you'll love -- pipeline-level quantization. Pass a quant config directly while `DiffusionPipeline.from_pretrained()` 🔥 The example above is the easiest entry point. But maybe you want more flexibility. For example, have different quantization configs for different components. Specify configs, not entire models. You retain all the flexibility, but with more ease 🤗 📜 Docs: https://lnkd.in/gMHKePEz 👨💻 PR: https://lnkd.in/gixRmKc6

Try Tencent's new Primitive-Anything app: https://lnkd.in/gG9zdH6j 🥇 First AI that breaks down complex 3D objects into simple geometric building blocks just like us humans do 🤙 Works on anything -- furniture to vehicles to characters 🎮 Perfect for game design, 3D modeling

I've made an open version of Google's NotebookLM, and it shows the superiority of the open source tech stack! 💪 The app's workflow is simple. Given a source PDF or URL, it extracts the content from it, then tasks Meta's Llama 3.3-70B, with writing the podcast script, with a good prompt crafted by Gabriel Chua ("two hosts, with lively discussion, fun notes, insightful question etc.") Then it hands off the text-to-speech conversion to Kokoro-82M, and there you go, you have two hosts discussion any article. The generation is nearly instant, because: > Llama 3.3 70B is running at 1,000 tokens/seconds with Cerebras Systems inference > The audio is generated in streaming mode by the tiny (yet powerful) Kokoro, generating voices faster than real-time. And the audio generation runs for free on Zero GPUs, hosted by HF on H200s. Overall, open source solutions rival the quality of closed-source solutions at close to no cost! Generate your podcast here: https://lnkd.in/eKiRRxGr

Nvidia dominating the top trending open datasets!

An improved version of the Segment Anything Model is now available in the Transformers library. Read more below:

Fine-tuning HiDream with LoRA has been challenging because of the memory constraints! But it's not right to let that come in the way of this MIT model's adaptation. So, we have shipped QLoRA support in our HiDream LoRA trainer 🔥 Don't let the sound of QLoRA scare you. The changes are way simpler than you think, thanks to PEFT and how it integrates with Diffusers. Here's the PR that landed this support: https://lnkd.in/gFumAkPn Without much bells and whistles, we go from ~60GB down to 37GB just by keeping the base model in NF4. massive, ain't it? I am sure there are other ways to squeeze out even more. But I purposefully didn't explore them. Find out the script below to learn more: https://lnkd.in/g_Y8ntSg H/T to Linoy Tsaban for the awesome support, especially the dataset and the validation prompt 🎬

Exciting Release! I've been building a new application called page-to-video that automatically turns web pages into long form video lessons with slides! 🎬 🔗 https://lnkd.in/ezzEVYbN If you’re sick of dense articles and want to digest information faster? page-to-video takes a URL and generates a video with audio and visual slides, making learning more accessible. Here's how it works: - You provide a webpage URL and page-to-video extracts the content. - It uses Cohere Labs LLMs to generate a transcription and editable markdown slides - fal text-to-speech model creates audio for each slide - Finally, it combines everything into a video. This experimental project showcases the power of combining different AI services through Inference Providers on Hugging Face   Check out the blog post to learn more about how it works.

Fine-tuning HiDream with LoRA has been challenging because of the memory constraints! But it's not right to let that come in the way of this MIT model's adaptation. So, we have shipped QLoRA support in our HiDream LoRA trainer 🔥 Don't let the sound of QLoRA scare you. The changes are way simpler than you think, thanks to PEFT and how it integrates with Diffusers. Here's the PR that landed this support: https://lnkd.in/gFumAkPn Without much bells and whistles, we go from ~60GB down to 37GB just by keeping the base model in NF4. massive, ain't it? I am sure there are other ways to squeeze out even more. But I purposefully didn't explore them. Find out the script below to learn more: https://lnkd.in/g_Y8ntSg H/T to Linoy Tsaban for the awesome support, especially the dataset and the validation prompt 🎬

Still hunting for robust LLM evaluation tools and I'm settled on using YourBench to create a series of detailed and high quality benchmarks that relate to your exact use case. I reckon the best way to use it as a set of configs within your project code base. You can get then use this to evaluate and review model, software, or prompt changes. This is what YourBench does: - build custom evaluation suites for your specific use cases or data examples. - - designed to work with the Hugging Face models and inference providers. - uses community backed datasets and evaluation methods - packaged with LLM performance across diverse tasks and metrics

Finally caught up on some long-overdue documentation — and sharing a few Hugging Face datasets I’m especially excited about from GLAM (galleries, libraries, archives, museums). 📚 Europeana Newspapers A massive multilingual corpus (~32B tokens!) of historical newspapers from the Europeana archive, curated for the BigLAM initiative. 🗣 12 languages 🧾 OCR confidence scores 🧠 Great for training/fine-tuning LLMs and digital history research 🔗 https://lnkd.in/e5PHVtW9 📰 Beyond Words A richly annotated visual dataset from WWI-era US newspapers (via Library of Congress Labs). 📦 3.5K pages 📸 48K+ bounding boxes 🔍 Visual categories: photos, maps, cartoons, headlines, ads 🧠 Used to train an Ultralytics YOLOv11 models for visual content detection 🔗 Dataset: https://lnkd.in/eVd5a8MS 🔗 Model: https://lnkd.in/eteJXwZe Excited to share more GLAM-focused AI work soon!