Brief-details: DiffRhythm-full is a groundbreaking diffusion-based AI model capable of generating complete songs, featuring a 4m45s generation capacity and latent diffusion architecture.
BRIEF DETAILS: 7.76B parameter multilingual LLM optimized for English, Korean, Japanese & Chinese. Strong performance in reasoning and instruction following, with 4K context window
Brief Details: LoRA model for generating mythical/fantasy artwork, specializing in creatures, ruins, and weapons. Requires 'cl4ud3' trigger. Non-commercial license.
BRIEF-DETAILS: Gemma 3 4B instruction-tuned multimodal model supporting text+image input with 128K context. Available in multiple quantization formats for various hardware configs.
BRIEF-DETAILS: A specialized LoRA model for Wan2.1 14B I2V that transforms images into gun-shooting videos, trained on diverse subjects with consistent results at 480p resolution.
Brief-details: A specialized LoRA model for Wan2.1 14B I2V that creates deflation effect animations, transforming objects into deflating versions with 20-epoch training on deflation videos
BRIEF-DETAILS: A LoRA model trained with Flux for image generation, requiring TOK trigger words. Built on Canopus-LoRA-Flux-UltraRealism-2.0 base model.
Brief Details: TraceBack-12b: A 12B parameter model built on Mistral Nemo, specialized in generating reasoning traces from instruction-solution pairs for synthetic dataset creation.
Brief-details: VRAM-16 is a memory optimization model by unslothai, designed for efficient GPU memory usage. Available on HuggingFace for memory-conscious AI applications.
Brief Details: A 4-bit quantized version of Vicuna-13B optimized for efficient local deployment using GPTQ compression with 128 group size, offering high-quality performance.
Brief-details: VGGT-1B is Facebook's 1B-parameter Transformer model for 3D scene understanding, capable of inferring camera parameters, depth maps, and point tracks efficiently.
Brief Details: IndicF5 - A polyglot TTS model supporting 11 Indian languages with near-human quality, trained on 1417 hours of speech data across major Indic languages.
Brief Details: Advanced 14B parameter LLM based on Qwen2.5, achieving top performance among sub-32B models. Features LoRA training and comprehensive instruction tuning.
Brief-details: Korean-optimized Sentence Transformer model based on Snowflake's arctic-embed, achieving SOTA performance in Korean retrieval tasks with 1024-dimensional embeddings.
Brief-details: VideoPainter is an advanced AI model for video inpainting and editing, featuring plug-and-play context control and any-length video processing capabilities, built on CogVideoX-5B architecture.
Brief Details: Ruri-large-v2 is a 337M-parameter Japanese text embedding model achieving 74.55% average score on JMTEB, with 1024-dimensional outputs and strong retrieval capabilities.
Brief Details: Cross-encoder model optimized for MS Marco passage ranking, offering excellent NDCG@10 score of 73.04 and processes 2500 docs/sec on V100 GPU.
Brief-details: A calibrated tiny T5 transformer model designed for testing and evaluation purposes, featuring improved probability calibration compared to standard T5 models.
BRIEF-DETAILS: Chronos-Bolt-tiny is a 9M parameter time series forecasting model based on T5 architecture, offering zero-shot predictions with 250x faster inference than original Chronos models.
Brief Details: Real-time face animation tool that maps webcam facial expressions to static images/videos. Uses TensorRT optimization & facial landmark detection for live performance.
Brief-details: Conversational Speech Model (CSM) with 1B parameters, optimized for MLX inference. Specializes in natural language processing with conversion to safetensors format.
