Quanta AI Labs

35 / Aug 25 – 31, 2025 · 12 RELEASES

AI News Week 35, 2025

Week 35 2025 AI releases.

Vision & Image (1)

  • NiT (Native-resolution diffusion Transformer)

    MMLab CUHK / Shanghai AI Lab · Text-to-Image

    A diffusion transformer architecture that generates images at arbitrary resolutions and aspect ratios using dynamic tokenization, variable-length sequence processing, and 2D structural priors. Achieves SOTA performance on ImageNet benchmarks (2.03 FID at 256×256, 1.45 FID at 512×512) with strong zero-shot generalization to unseen resolutions up to 2048×2048.

Video & Animation (6)

  • FlowMo

    Tel Aviv University (Ariel Shaulov, Itay Hazan, Lior Wolf, Hila Chefer) · Text-to-Video

    A training-free guidance method that improves motion coherence in text-to-video diffusion models by measuring patch-wise variance across latent frames and dynamically reducing it during sampling. Works as a plug-and-play solution on pre-trained models without retraining or auxiliary inputs.

  • Ctrl-Crash

    Mila - Quebec AI Institute · Text-to-Video

    A controllable video diffusion model that generates realistic car crash scenarios conditioned on bounding boxes, crash types, and initial frames, using classifier-free guidance for fine-grained control over multiple conditioning signals. Achieves state-of-the-art performance on video quality metrics and enables counterfactual accident simulation for traffic safety research.

  • DeepVerse

    Shanghai AI Lab et al. · Text-to-Video

    A 4D autoregressive video generation model that uses textual prompts as universal control signals and incorporates depth information to maintain temporal coherence and enable physics-aware world modeling for extended video synthesis.

  • Bing Video Creator

    Microsoft · Text-to-Video

    Free AI video generation tool powered by Sora that converts text prompts into 5-second videos. Rolling out on Bing Mobile App with desktop and Copilot Search integration, offering Fast and Standard generation speeds with Microsoft Rewards redemption option.

  • SkyReels-Audio

    Skywork AI · Text-to-Video

    A unified video diffusion transformer framework for generating high-fidelity, temporally coherent audio-conditioned talking portrait videos with multimodal control (text, images, audio). Uses hybrid curriculum learning, facial mask loss, and sliding-window denoising to achieve superior lip-sync accuracy and identity consistency across infinite-length sequences.

  • HunyuanCustom

    Tencent Hunyuan · Text-to-Video

    Multi-modal customized video generation framework supporting image, audio, video, and text conditions with enhanced subject identity consistency through text-image fusion, temporal ID enhancement, and modality-specific injection mechanisms.

Audio & Speech (1)

  • OpenAudio S1

    OpenAudio (Hanabi AI Inc.) · Text-to-Speech

    A 4B-parameter text-to-speech model trained on 2M+ hours of audio that achieves state-of-the-art WER/CER and ranked #1 on HuggingFace TTS-Arena-V2, featuring precise emotional and tonal control through instruction markers and native support for 13 languages at $0.8/hour.

3D & Spatial (2)

  • ShapeLLM-Omni

    Tsinghua University, Peking University, ShengShu · 3D Generation

    A native multimodal LLM extending Qwen-2.5-vl to understand and generate 3D assets alongside text and images, using a 3D VQVAE for discrete shape representation and the new 3D-Alpaca dataset for instruction-based training on generation, comprehension, and editing tasks.

  • Pixel3DMM

    Technical University of Munich, Synthesia, University College London · 3D Reconstruction

    A fine-tuned DINO ViT that predicts per-pixel surface normals and UV coordinates to constrain 3D morphable face model optimization for single-image 3D face reconstruction. Achieves 15% improvement over baselines on a new benchmark evaluating both posed and neutral facial geometry across diverse expressions, angles, and ethnicities.

Developer Tools (2)

  • ComfyUI-Lora-Manager

    willmiao · AI Infrastructure

    A manager extension for ComfyUI that simplifies LoRA (Low-Rank Adaptation) model management, enabling users to organize, load, and apply LoRA weights within the ComfyUI node-based interface.

  • Google AI Edge Gallery

    Google AI Edge · Edge Deployment

    A curated collection of example applications and reference implementations demonstrating on-device ML capabilities using TensorFlow Lite and MediaPipe frameworks for edge deployment.

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