Quanta AI Labs

09 / Feb 24 – Mar 2, 2025 · 10 RELEASES

AI News Week 9, 2025

Week 9 2025 AI releases.

Language Models (2)

  • CSM-1B

    Sesame · Small Language Models

    A 1 billion parameter model available on Hugging Face. Specific capabilities and architecture details are unavailable due to page truncation.

  • Gemma 3

    Google · Reasoning

    Gemma 3 is Google's latest open-source language model series offering improved performance across reasoning, coding, and multimodal capabilities compared to previous versions.

Vision & Image (2)

  • SANA-Sprint

    NVIDIA · Text-to-Image

    One-step diffusion model for ultra-fast text-to-image generation achieving 7.59 FID in 0.1s on H100, 10× faster than FLUX-schnell through continuous-time consistency distillation and hybrid training-free transformation of pre-trained flow-matching models.

  • Thera Arbitrary-Scale Super-Resolution

    prs-eth · Image Restoration

    Image super-resolution model capable of upscaling images to arbitrary scales. Deployed as an interactive Hugging Face Space for real-time inference.

Video & Animation (3)

  • WAN21 14B 480p I2V LoRAs

    Remade-AI · Text-to-Video

    A collection of LoRA (Low-Rank Adaptation) fine-tuned models for the WAN21 14B image-to-video model, optimized for 480p resolution video generation. These adapters enable customized video generation capabilities with reduced computational overhead.

  • Long Context Tuning (LCT) for Video Generation

    ByteDance Seed / Chinese University of Hong Kong · Text-to-Video

    LCT extends video diffusion models from single-shot to scene-level generation by expanding MMDiT's context window using interleaved 3D RoPE and asynchronized diffusion timesteps, enabling coherent multi-shot narratives with emergent capabilities like interactive extension, compositional generation, and scene interpolation.

  • ReCamMaster

    Zhejiang University, Kling Team (Kuaishou Technology), CUHK, HUST · Video Editing

    ReCamMaster enables camera-controlled generative re-rendering of videos by applying novel camera trajectories to source footage through a video conditioning mechanism built on pre-trained text-to-video diffusion models. The framework supports diverse applications including video stabilization, 4D reconstruction, autonomous driving data augmentation, and embodied AI, with a newly released multi-camera synchronized dataset from Unreal Engine 5.

3D & Spatial (3)

  • MeshPad

    Technical University of Munich & AUDI AG · 3D Generation

    MeshPad is a sketch-conditioned generative approach for interactive 3D mesh creation and editing that decomposes the task into addition and deletion operations, using a Transformer-based triangle sequence model with vertex-aligned speculative prediction for real-time inference (few seconds per edit).

  • MIDI-3D

    VAST-AI · 3D Generation

    A generative model that converts MIDI music files into 3D visual representations. Combines audio-to-visual synthesis with 3D rendering capabilities.

  • VGGT: Visual Geometry Grounded Transformer

    Visual Geometry Group (University of Oxford) & Meta AI · 3D Reconstruction

    A feed-forward transformer that directly infers 3D scene attributes including camera parameters, depth maps, point clouds, and tracks from single or multiple images in under one second. Achieves state-of-the-art results across camera estimation, multi-view depth, dense reconstruction, and point tracking tasks.

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.