Real-ESRGAN is an open-source AI model that dramatically enhances low-resolution images using blind super-resolution and synthetic degradation modeling—no high-res reference needed.
DemoFusion is an open-source AI framework that generates ultra-high-resolution images (up to 4096×4096+) from low-res inputs using progressive upscaling—no extra training or expensive hardware required.
Draw an Audio is a research-grade AI system that automatically generates contextually accurate, temporally aligned sound effects from silent videos using multimodal conditioning.
ActAnywhere is a research video generation model that automatically creates temporally consistent, motion-aware backgrounds for foreground subjects—enabling seamless background replacement and visual effects without manual keyframing or green screens.
FineVideo is a large-scale, open multimodal video dataset designed to advance AI models’ understanding of spatial-temporal scenes, emotions, narratives, and cross-modal relationships in video.
Vary-toy is a compact, open-source visual language model enabling document OCR, image description, visual question answering, and multimodal dialogue on consumer-grade hardware.
Follow Your Pose is an open research framework that generates highly controllable, temporally coherent videos from text prompts and user-specified human pose sequences.
IP-Adapter is an open-source adapter that lets you guide Stable Diffusion and similar models using image prompts—enabling accurate, controllable image generation without fine-tuning.
OLMo is a fully open, high-performance language model framework offering transparent training data, reproducible code, and standardized evaluation tools for researchers and developers.
DiffusionGPT is an open-source, LLM-driven text-to-image system that intelligently selects and coordinates specialized diffusion models to generate high-fidelity images from diverse, complex prompts.
SafeEar is an open-source AI framework that detects audio deepfakes by analyzing acoustic features while preserving speaker privacy across multiple languages.
MotionCtrl is an open-source motion controller for video generation models that enables independent, precise control of camera movement and object motion in AI-generated videos.
Llama 3.2 is Meta's open-weight multimodal model family optimized for edge devices, supporting vision-language understanding, on-device text generation, and customizable local deployment.
MMMLU is a large-scale, multilingual, multitask benchmark dataset for rigorously evaluating language models across 14 languages and 57 academic disciplines.
AnimateDiff is an open-source framework that transforms existing text-to-image models into animation generators without retraining, enabling high-fidelity, style-consistent video sequences from text prompts.
PortraitGen is an open-source AI video editing tool that enables high-fidelity, temporally consistent editing of human portrait videos using 3D Gaussian splatting and neural texture modeling.
SFR-RAG is an open-source, retrieval-augmented language model from Salesforce AI Research that improves factual accuracy, reduces hallucinations, and supports reliable, context-grounded text generation and conversational AI.
V-JEPA is an open-source, self-supervised video foundation model that learns visual representations by predicting masked frame features—enabling robust video understanding without labels.