Hugging Face

Logo in Context -- Apply any logo to anything!?? Cool trick to copy a logo to create useful mockups or just for fun with your friends! ?? Workflow: In-Context LoRA + Image-to-Image + Inpainting. Explore the free gradio app and learn more about the project here on Hugging Face Spaces: https://lnkd.in/ga2NCtQ6

Lets go! ?? Generate images based on the reference image and a text prompt. InstantX Team has released the first IP Adapter for FLUX [dev]. This has been a tremendous release that very nicely complements the recently launched FLUX.1 Tools by Black Forest Labs. The InstantX/FLUX 1-dev IP Adapter Gradio app is available for FREE for the community to explore. You can access it on Hugging Face Spaces at: https://lnkd.in/gr5AUGFD.

?? Introducing EchoMimicV2 for simple semi-body human animations! ??It utilizes a reference image, an audio clip, and a sequence of hand poses to generate a high-quality animation video, with coherent audio and half-body movements. License: Apache 2.0. Deploy EchoMimicV2 locally with the official Gradio app: https://lnkd.in/g_fqXf43 Model weights on Hugging Face: https://lnkd.in/gyJxNMNm

Having preset options that a user can click on in a chat!?? ?? ?? Coming soon to Gradio chatbots!

What a week! A recap for everything you missed ?? Link to all models and datasets in comments ?? Multimodal ? > Mistral AI released Pixtral 124B, a gigantic open vision language model > Llava-CoT (formerly known as Llava-o1) was released, a multimodal reproduction of o1 model by PKU > OpenGVLab released MMPR: a new multimodal reasoning dataset > Jina has released Jina-CLIP-v2 0.98B multilingual multimodal embeddings > Apple released new vision encoders LLMs ?? > Ai2 dropped a huge release of models, datasets and scripts for Tülu, a family of models based on Llama 3.1 aligned with SFT, DPO and a new technique they have developed called RLVR > Jina has released embeddings-v3: new multilingual embeddings with longer context > Hugging Face released SmolTalk: synthetic dataset used to align SmolLM2 using supervised fine-tuning > Microsoft released orca-agentinstruct-1M-v1: a gigantic instruction dataset of 1M synthetic instruction pairs Image Generation ??? > Black Forest Labs released Flux 1. tools: four new models for different image modifications and two LoRAs to do image conditioning and better steer generations Lastly Hugging Face released a new library Observers: a lightweight SDK for monitoring interactions with AI APIs and easily store and browse them ?? $ pip install observers

???? If you think transparency is key to building the image generation models of tomorrow, consider contributing to a massive open dataset. What do you need to do: 1 Follow data-is-better-together on Hugging Face right now 2 Follow me on LinkedIn and Bluesky 3 Have a crazy weekend 4 Start the annotation sprint with Hugging Face and Argilla on Monday 25th

Ai2 Tülu 3, my favourite Open Source AI release ever! As Nathan Lambert says, it sets the next era in open post-training. My highlight? The data generation pipelines and the open datasets Want to dive deep into the data? Here's a Hugging Face Space to understand the data powering this open frontier model. https://lnkd.in/d-YRtXqZ

?? CrisperWhisper is an advanced variant of OpenAI's Whisper, designed for fast, precise, and verbatim STT with crisp word-level timestamps! ?? Unlike the original Whisper, nyra health's CrisperWhisper aims to transcribe every spoken word exactly as it is, including fillers, pauses, stutters and false starts. CrisperWhisper is on top of Open ASR Leaderboard??on Hugging Face Spaces: https://lnkd.in/gSwuKvZB Key Features ?? Accurate Word-Level Timestamps: Provides precise timestamps, even around disfluencies and pauses, by utilizing an adjusted tokenizer and a custom attention loss during training. ?? Verbatim Transcription: Transcribes every spoken word exactly as it is, including and differentiating fillers like "um" and "uh". ?? Filler Detection: Detects and accurately transcribes fillers. ??? Hallucination Mitigation: Minimizes transcription hallucinations to enhance accuracy.

Today is my first day at Hugging Face! I’m extremely happy to share that I’m joining the incredible Open-Source Team as a full time ML enginner to work on Transformers with Arthur Zucker, Jo?o Gante, and many others! This opportunity came after I submitted two PRs with huge impact on memory savings during inference. The first one (https://lnkd.in/eDhKTXw8) addressed a long time issue I noticed, in which the K-V cache was duplicated at inference time, leading to twice the necessary memory usage. Thus, it decreased the memory usage by a factor 2 (3 for beam-based decoding strategies) for all recent enough causal models. This one was already integrated in Transformers v4.42. In the 2nd one (https://lnkd.in/eACAVssk), I allowed to only compute the necessary logits (in contrast with the full logit tensor), in the first forward pass when performing iterative decoding. As the vocabulary size of the models gets bigger and bigger, the full logit tensor becomes enormous as the input size increases, but we never need it all at inference time! The result is massive memory gains for recent models: up to a factor 10 for Gemma 2 with its 256K vocabulary size, and a factor of 3.6 for the Llama 3 family (both 3.1 and 3). To see such factors in practice, you need the input size to be bigger than the number of tokens you generate. But even if this condition is not met, you will still benefit from massive savings (just slightly lower than the previously discussed ratios). This one was already merged, but is not yet released! It will be live with Transformers v4.45 in a few days! That’s great, but at what cost? Absolutely none. There are literally 0 performance penalties to these optimizations. It’s free lunch for the whole community. Figures 1 and 2 show the impact of the 2nd PR on Gemma 2 and Llama 3.1, while Figure 3 show the impact of the 1st one on Llama 3 (Llama 3.1 and Gemma 2 did not exist yet at the time!). Everything was benchmarked with a RTX 4090 (24 GB of ram).

Let's go! Hyperbolic Labs has added a gradio integration to their Playground, which also allows you to deploy to Hugging Face Spaces in just one click!??? ?? Select any LLM and deploy a chat interface on Spaces to begin interacting with the model. It's a great way to quickly try out different LLMs and serves as a good starting point for building more complex apps. One-click app deployment (within seconds) on Hyperbolic labs AI Playground!?? ??Try it out yourself for Qwen2.5 Coder 32B model: https://lnkd.in/gMcWaGmY