This Week in AI: GPT-5.2, Wan-Move, AutoGLM & More

TL;DR

This week's AI releases include GPT-5.2 from OpenAI, Alibaba's Wan-Move for controlling video motion by drawing trajectories, and AutoGLM, an open-source 9-billion-parameter agent that autonomously operates phones. Other tools remove window reflections from photos, generate hyper-realistic images, and create editable multi-part 3D models from reference images.

Transcript

AI never sleeps and this week has been absolutely insane. Open AAI drops their latest model GPT 5.2. Alibaba releases a very useful video tool that allows you to control the motion of objects in the video by drawing trajectories. We have an AI that can edit videos in real time. You can insert or replace anything with just a prompt. This AI can crea... Read More

Key Insights

• Window Seat is an AI reflection-removal tool that magically cleans up window reflections in photos, correcting brightness, contrast, and colors, and outperforms alternatives like DSIT, RDNet, and DAI in quantitative benchmarks across different lighting conditions.
• Window Seat is a LoRA for Qwen Image Edit, and while the GitHub repo lists a 24 GB VRAM CUDA GPU requirement, quantized compressed versions let you run it comfortably on low VRAM.
• RealGen is a realistic image generator that uses a detector reward mechanism, rewarding the model when generated images avoid flaws like plasticky looks or overly smooth skin, making it more realistic than Flux, Qwen Image, and Flux Krea.
• Wan-Move from Alibaba lets you control object motion in video by drawing trajectories on the start frame, supports multiple trajectories at once plus camera movements, and outperforms the proprietary Kling 1.5 Pro in physical consistency.
• AutoGLM from ZAI is an open-source AI agent that autonomously operates your phone, navigating apps, sending messages, and even making payments without human intervention, such as finding cinemas on Google Maps or summarizing X posts.
• AutoGLM is only 9 billion parameters and roughly 18 GB in size, too large for most smartphones but able to fit on a consumer-grade GPU running an Android simulator.
• MoCA is a 3D model generator that creates complex objects and scenes from a reference image while separating the models into meaningful parts, so each piece can be edited, exploded, or animated by choosing joints.
• Wan-Move's full model is nearly 70 GB, but a Comfy UI integration by Kijai offers an FP8 compressed version at only 16 GB that can fit within 16 GB VRAM with some offloading.

Questions & Answers

Q: How do you remove window reflections from photos using AI?

Window Seat is a new AI tool that magically removes reflections from photos taken through glass, such as on a plane, at a zoo, or through rainy windows. It even corrects brightness, contrast, and colors. It is a LoRA for Qwen Image Edit, and although the GitHub repo lists a 24 GB VRAM CUDA GPU requirement, quantized compressed versions of Qwen Image Edit let you run it comfortably on low VRAM. It outperforms alternatives like DSIT, RDNet, and DAI in benchmark scores.

Q: What is Wan-Move and how does it control video motion?

Wan-Move is a new AI tool from Alibaba that lets you control the motion of objects in a video simply by drawing trajectories on the start frame. You draw how you want things to move, and the resulting video follows. It supports multiple trajectories at once, follows the laws of physics such as pouring liquid from a tilted kettle, and can also control entire camera movements like panning, dolly in or out, and zooming. It outperforms the proprietary Kling 1.5 Pro in physical consistency.

Q: What is AutoGLM and what can it do on a phone?

AutoGLM, released by ZAI (the company behind the open-source GLM models), is an open-source AI agent that autonomously operates your phone. It can navigate through apps, send messages, and even make payments. Examples include finding top-rated cinemas nearby and navigating there on foot via Google Maps, liking and summarizing the top three posts from a user on X, writing and sending emails, and shopping by adding running shoes to a cart and going to checkout, all without human intervention.

Q: How many parameters does AutoGLM have and can I run it locally?

AutoGLM is only 9 billion parameters, making it fairly small to download. According to the Hugging Face link, the model is roughly 18 GB in size. While it might not fit on most smartphones yet, it can comfortably fit on a consumer-grade GPU on your computer, where you can run an Android simulator to test it out. It is open source like ZAI's other models, and download links plus technical details are available on the project page.

Q: What is MoCA and how does it generate 3D models?

MoCA is a new 3D model generator that is good for creating complex 3D objects and scenes from a reference image. Beyond generating the model, it separates the 3D models into meaningful parts so you can edit them afterwards. For example, inputting an image of a dinosaur produces a detailed 3D model capturing markings, which can be exploded into parts. This is useful because you can edit each part further or animate the model by choosing the joints where it should move.

Q: How does RealGen generate realistic images?

RealGen is an image generator designed to produce super realistic images. It uses a detector reward mechanism that helps the model figure out how realistic a generated image is and make adjustments to improve it. The detectors check for artifacts or unrealistic features, such as plasticky looks or overly smooth skin, and reward the model when it avoids these flaws. After this training, RealGen outperforms other models like Flux, Flux Context, Qwen Image, and Flux Krea in realism, even adding photographic grain and motion blur.

Q: How can I run Wan-Move if I don't have a large GPU?

The official Wan-Move model file is around almost 70 GB in size, which is hard to fit with most consumer GPUs. Fortunately, there is already an integration for Comfy UI, and a developer called Kijai released an FP8 version that is more compressed at only 16 GB. This can likely fit within 16 GB of VRAM with some offloading. All the links, including the GitHub repo with instructions on downloading and running it locally, are available on the project page.

Q: What new AI models were announced this week?

This week's releases included OpenAI's GPT-5.2, Alibaba's Wan-Move video motion tool, and ZAI's AutoGLM phone agent. Other releases covered a real-time video editor that inserts or replaces anything with a prompt, an AI creating 3D stereo videos, Window Seat for removing window reflections, RealGen and a fast image model generating images in about a second, an anime-focused lightweight image model, an open-source character animation model, and MoCA, a top 3D model generator making objects in separate parts.

Summary & Key Takeaways

• The week featured major AI releases including OpenAI's GPT-5.2, a real-time video editor, AI stereo video generation, mobile AI agents, and full body control tools. The video walks through each release with examples, benchmarks, and links to GitHub repos for running them locally.

• Standout image tools include Window Seat, which removes annoying window reflections from photos taken through glass on planes, at zoos, or in the rain, and RealGen, which uses a detector reward mechanism to produce hyper-realistic images with grain and motion blur effects.

• Video and 3D advances include Alibaba's Wan-Move for trajectory-based motion and camera control that beats Kling 1.5 Pro, ZAI's open-source AutoGLM phone agent at 9 billion parameters, and MoCA, which builds detailed, editable multi-part 3D models from a single reference image.