Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 5, 2026, 9:00 AM
Hugging Face's official blog summarizes the progress of its computer vision ecosystem, covering 8 core visual tasks (image classification, zero-shot object detection, video classification, etc.), over 3000 models, and 100+ datasets.
The Hugging Face blog published the 4th article in the "AI for Game Development" series, detailing how to integrate Stable Diffusion as an auxiliary tool into traditional 2D asset workflows.
Hugging Face has officially integrated LoRA support into the diffusers library for efficient fine-tuning of Stable Diffusion and Dreambooth. This feature allows running full fine-tuning scripts on 11 GB VRAM (e.g., RTX 2080 Ti). The trained LoRA weights are only about 3 MB and can be stacked onto the original model for inference; documentation, example commands, Hub API usage, and Dreambooth integration details have all been made public.
The Hugging Face Optimum library has officially integrated support for ONNX Runtime training, achieving up to 130% faster training speeds (compared to the PyTorch baseline) for various Hugging Face models. It supports single-node setups with 8x A100 GPUs, offers an ORTTrainer API to replace the original Trainer, is compatible with DeepSpeed ZeRO-1 and mixed-precision training, and supports model saving and uploading to the Hub.
The official Hugging Face blog announced an open-source collaboration with Baidu PaddlePaddle, fully integrating PaddlePaddle models into the Hugging Face Hub.
This article introduces how to build an image similarity system using 🤗 Transformers and 🤗 Datasets. Using the ViT model nateraw/vit-base-beans and the Beans dataset as examples, it demonstrates the full process of image embedding extraction, cosine similarity calculation, and top-k retrieval. The text also mentions methods for FAISS integration and provides a link to the accompanying Colab Notebook.
Hugging Face released the first part of a tutorial series, demonstrating how to generate isometric-style farm concept art using Stable Diffusion 1.5 (prompts include "isometric render").
Hugging Face released a tutorial introducing the CLIPSeg zero-shot image segmentation model, which supports text prompts (e.g., 'pancakes') and visual prompts (using example images as prompts) to output coarse-grained segmentation masks. The article provides complete 🤗 transformers invocation code, visualization methods, and demonstrates how to upload results to Segments.ai for manual refinement.
Hugging Face has released a creation tool for Model Cards and new templates, including a graphical interface tool, updated huggingface_hub library templates, annotated templates, a user research report, and a literature review; this work is based on the concept of Model Cards proposed by Mitchell et al. in 2018, aiming to improve the accessibility, inclusivity, and standardization of machine learning model documentation.
Hugging Face's official blog published "A Complete Guide to Audio Datasets," systematically introducing how to use the 🤗 Datasets library to download and preprocess audio datasets.
Hugging Face has released the Bumblebee library, bringing models such as GPT2 and Stable Diffusion into the Elixir ecosystem; this library is a pure Elixir implementation of Transformers.
The Hugging Face blog announced a collaboration with Apple to convert official checkpoints such as Stable Diffusion v1.4/v1.5/v2-base/v2.1-base into Core ML format and host them on the Hugging Face Hub.
OpenAI released a model named ChatGPT, which supports conversational interactions, can answer follow-up questions, admit mistakes, challenge incorrect premises, and reject inappropriate requests.
Hugging Face Spaces has integrated with arXiv, adding a Demo tab to each arXiv paper to showcase open-source machine learning demos created by the community or authors; since its launch in October 2021, Spaces has hosted over 12.
Hugging Face blog published a tutorial using the Concrete-ML library to build a sentiment analysis model in a Fully Homomorphic Encryption (FHE) environment: extracting text representations with cardiffnlp/twitter-roberta-base-sentiment-latest.
The official Hugging Face blog introduced Contrastive Search, a new decoding method proposed by Su et al. in papers at NeurIPS 2022 and arXiv 2022.
Hugging Face announced the discontinuation of the paid tier for the Inference API, which remains available for free; simultaneously, it launched enterprise-grade managed inference service Inference Endpoints and upgraded hardware support for Spaces; all paid services are billed monthly with consolidated invoices, and users need to link a credit card but do not require a subscription to use them.
The Hugging Face blog published a practical guide for training Stable Diffusion using the Diffusers library, summarizing learning rates and training steps.
Developers can now begin using the DALL·E API to build applications.
The Hugging Face official blog published a complete tutorial on using 🤗 Transformers to fine-tune the Whisper model for multilingual automatic speech recognition (ASR), using Hindi as an example.
Hugging Face announced the integration of Intel OpenVINO 2022.2 into Optimum Intel, supporting the execution of Transformers models on various Intel processors and performing post-training static quantization.
PyTorch 1.13 has been officially released, featuring the stable version of BetterTransformer (supporting inference acceleration for Transformer models), the formal integration of the functorch library into the PyTorch main branch (no separate installation required), and native support for Apple M1 chips entering the Beta stage. Additionally, CUDA 10.2 and 11.3 have been deprecated, with migration completed to CUDA 11.6/11.7.
In a blog post, Hugging Face introduced how to use the 🤗 Evaluate library to assess biases in large language models (such as GPT-2 and BLOOM) across three types of tasks: Toxicity, Polarity, and Hurtfulness. It provided dataset invocation examples for WinoBias, BOLD, and HONEST, along with complete Python code, supporting toxicity ratio calculation, Regard difference analysis, and HONEST gender stereotype bias scoring.
The official Hugging Face blog published a tutorial systematically explaining three abstraction levels for implementing distributed training in PyTorch: native torch.distributed DDP, the 🤗 Accelerate lightweight wrapper, and the 🤗 Transformers high-level Trainer API. Using an MNIST classification task as an example, it provides complete runnable code that supports single GPU, multi-GPU, and TPU without requiring modifications to the core logic.