Real gundams, top 3D generator, open-source world models, ChatGPT updates, new TTS: AI NEWS

JDJust Dub ItAvailable in EU

Video dubbing AI that translates speech into another language and adjusts lip movements. Built as a small 2.5 GB LoRA on LTX 2.3, reported to outperform methods like HeyGen. Code and model released.

PIPixal3DUnverified

Open-source image-to-3D generator using pixel-aligned generation for accurate geometry and texture. Beats Hunyuan 3D and Trellis 2 in comparisons; models total 24 GB.

ASAsymFlowUnverified

Image generator that works directly in pixel space without a VAE or latent space. Architecture tweaks made it 40% faster and it gives more photorealistic results; code and training code released.

SWSANA WMNVIDIAUnverified

2.8B-parameter open-source interactive world model that turns an image, prompt and WASD input into video on one GPU. A distilled variant makes a minute-long clip in 34 seconds on an RTX 5090; weights are coming soon.

WAWarp as HistoryUnverified

Interactive world generator taking a first frame, text prompt and keyboard input to generate consistent video. Code, models and training scripts released.

PHPhyMotionUnverified

Reward system that recovers a 3D human body from generated video and scores it in the MuJoCo physics simulator. This improves anatomical correctness in video models, though some errors remain. Code released.

IMInteraction ModelsThinking MachinesUnverified

Real-time models that take in audio, video and text continuously while thinking and responding, handling interruptions and visual cues. A lightweight interaction model hands off to a background model for deeper reasoning. A limited research preview is planned in the coming months.

CACausalCineUnverified

Video generation system that streams long multi-shot videos in real time from shot-level prompts, reusing memory for consistency. Outperforms Infinity Rope, LongLive, MemFlow and Self-Forcing; only a paper is out so far.

RERelit-LiVEUnverified

Video relighting tool that changes light warmth, hardness, angle and can project environment maps. Code released; needs a CUDA GPU with 24 GB VRAM.

M4MiniCPM-V 4.6OpenBMBAvailable in EU

Tiny vision-language model that runs on phones across iOS, Android and Harmony, with a 2.6 GB total size. It outperforms similar-sized models on document, chart, GUI and video understanding and is token-efficient.

K2Krea 2KreaAvailable in EU

Krea's first from-scratch image foundation model, focused on artistic style and creative control, with style-reference uploads. It is closed source and the creator thinks it trails GPT Image 2.

CICodex in ChatGPT mobile appOpenAIAvailable in EU

Lets users monitor, approve, redirect and start Codex tasks from the ChatGPT mobile app on iOS and Android, while files stay on the computer. Rolling out in preview on all plans; currently only works with Codex on macOS.

Open-source music generator that makes full songs from a text prompt and lyrics. Output sounds clean and stereo but has some lyric mispronunciations, and the creator isn't sure it beats ACE-Step 1.5 XL. Needs 24 GB VRAM.

Second-generation dexterous robotic hand with 23 degrees of freedom, 0.1 mm repeatability, 12 kg grasp load and slip detection. Tendon-driven design keeps it under 400 g.

UGUnitree GD01UnitreeUnverified

Manned transformable mecha robot that can also walk autonomously, weighing about 500 kg with an operator and starting at $650,000. It can walk bipedally or switch to quadruped movement and knock down a block wall.

ARArticraftUnverified

Generates articulated 3D assets with parts and joints by having a coding agent such as GPT-5.5 write a program that builds the object. Comes with the Articraft-10K dataset of over 10,000 assets in 245 categories; code released and agent-agnostic.

CPChatGPT Personal FinanceOpenAINot in EU

Personal finance experience in ChatGPT that connects accounts via Plaid, builds dashboards and answers questions using real financial data. It can't see full account numbers or change accounts. Preview for Pro users in the US on web and iOS.

TRTrackCraft3rAvailable in EU

Repurposes a video diffusion model to track every pixel in 3D through a video. It beats Motion Tracker and Any4D on benchmarks, handles long videos and large motion, and uses less memory. Code released.

MOMoCamUnverified

Changes the camera movement of an existing video while preserving subject motion, including bullet-time effects. Code and models are coming soon.

DWDreamX WorldAlibaba (Amap)Available in EU

Interactive world generator with keyboard control and mid-generation prompts, supporting first and third person views and varied styles. Only the 5B-parameter model is released so far; larger and real-time variants are planned.

GAGemini AI pointerGoogle DeepMindUnverified

Prototype Gemini-powered mouse pointer that understands what you point at, so you can say "summarize this" or make a chart from a table. It is being connected to Chrome and Google Book.

SAScenema AudioAvailable in EU

Expressive open-source TTS built from the audio model of LTX 2.3. It clones voices, follows stage directions and emotion, changes accents, supports multiple languages and designs voices from text prompts. Needs 16 GB VRAM with a quantized model plus 32 GB RAM.

DRDramaBoxResemble AIAvailable in EU

TTS extracted from LTX 2.3 that clones voices, takes scene direction and produces vocalizations like laughs and sighs. It needs a 7 GB audio model plus an 8 GB Gemma text encoder, with 24 GB peak VRAM.