Tag

An open-source repository from Stability AI that collects implementations, training configs, demos and inference scripts for multiple generative models (e.g. SDXL, SV3D, SV4D, SV4D 2.0, Stable Video Diffusion). It is modular and config-driven, provides sampling/demo scripts, training examples, and references to model weights on Hugging Face.

Deep-Live-Cam is an open-source real-time face-swap / deepfake tool that can replace faces in live webcam streams or videos using only a single source image. Key features include one-click live deepfakes, mouth-mask to retain original mouth motion, multi-face mapping, and pre-built binaries for Windows and Apple Silicon. It supports multiple execution providers (CUDA, CoreML, DirectML, OpenVINO) and includes built-in content checks and ethical guidance.

screenpipe is an open-source AI app store that powers desktop-native AI apps using continuous (24/7) local desktop history (screen + mic recording). It is 100% local, developer-friendly, supports a plugin system for building and publishing apps, and includes features like native OCR and indexed desktop history for richer AI context.

VideoCaptioner is an AI-powered video subtitling assistant that combines ASR (local or cloud) with LLM-based subtitle segmentation, correction and translation. It supports offline GPU transcription, concurrent chunk transcription, VAD, speaker-aware processing, batch subtitling and one-click subtitle-to-video synthesis, with both GUI and CLI options.

ComfyUI-LTXVideo is a collection of custom nodes and workflows for ComfyUI that extend support for the LTX-2 video generation model. It makes LTX-2 features accessible inside ComfyUI, provides example pipelines (text-to-video, image-to-video, video-to-video), and includes instructions for required model files, upscalers, and low-VRAM usage.

VideoRAG (with the Vimo desktop app) is an open-source framework and application for chatting with videos. It uses retrieval-augmented generation and graph-driven knowledge indexing to enable understanding and question-answering over extremely long videos. Features include hierarchical context encoding, multi-modal retrieval (visual + audio + text), cross-video reasoning, a LongerVideos benchmark, and a desktop frontend for interactive video conversations.

ComfyUI-WanVideoWrapper is a collection of custom ComfyUI nodes that wrap WanVideo (Wan2.1 and related models) and related components, enabling easier loading and use of WanVideo-based video generation models inside ComfyUI. It supports multiple model formats, memory/VRAM optimizations, and provides example workflows and compatibility patches.

Memvid is a single-file, portable memory layer for AI agents. It packages content, embeddings, indices, and metadata into an append-only .mv2 file organized as immutable "Smart Frames", providing fast local retrieval, long-term persistence, time-travel debugging, and serverless operation without external vector DBs.

SAM 3 (Segment Anything with Concepts) is Meta's open-vocabulary foundation model for promptable image and video segmentation and tracking. It supports text and visual prompts (points, boxes, masks, exemplars) to detect, segment and track all instances matching a concept. SAM 3 introduces a presence token, a decoupled detector–tracker architecture, and is trained with a massive automatically annotated concept dataset (SA-Co). The GitHub repository provides code for inference, finetuning, example notebooks, and links to model checkpoints (Hugging Face access required).

This paper demonstrates the zero-shot learning and reasoning abilities of the generative video model Veo 3, paralleling the evolution of Large Language Models (LLMs) in natural language processing. Veo 3 excels in diverse visual tasks without explicit training, such as object segmentation, edge detection, image editing, understanding physical properties, recognizing affordances, and simulating tool use, enabling early visual reasoning like maze solving and symmetry detection.

LightX2V is an advanced lightweight video generation inference framework engineered to deliver efficient, high-performance video synthesis solutions. This unified platform integrates multiple state-of-the-art video generation techniques, supporting diverse generation tasks including text-to-video (T2V) and image-to-video (I2V). X2V represents the transformation of different input modalities (X, such as text or images) into video output (V).

The best introduction to how large language models (LLMs) like ChatGPT works in the world. It covers the three main stages of their training: pre-training on vast amounts of internet text, supervised fine-tuning to become helpful assistants, and reinforcement learning to improve problem-solving skills. The video also discusses LLM psychology, including why they hallucinate, how they use tools, and their limitations. Finally, it looks at future capabilities like multimodality and agent-like behavior.