Weight space programming, science agents, 3D generators, new Gemini Omni, new GPT Image: AI NEWS
The week’s most practical AI developments span faster image generation, 3D reconstruction, robotics, scientific agents, and weight-space model debugging.
- Try Nitro Fusion 2 for fast image generation; it can produce images in one or two steps and runs in ComfyUI or Diffusers.
- Use Hair3D, Flow HMR, or Cola 3D to turn images or video into 3D hair, motion, or object-separated scenes.
- Treat weight-space programming as an early debugging approach, not a proven way to upgrade any model: its gains depended on testing and a specific model flaw.
A broad news roundup; the notes capture the main launches and caveats without needing the full video. · Best for: AI practitioners tracking new models, open-source tools, robotics, and 3D generation.
A roundup of AI releases and research, ranging from image and video tools to 3D, robotics, scientific agents, and large language models. The most substantive research discussion examines weight-space programming and explains why its headline improvement may be overstated.
3D hair, video upscaling, and fast image generation1:06
- Hair3D reconstructs animated 3D hair from one image; its model and code are available, with support for Nvidia GPUs and Apple silicon.
- Relay VSSR upscales video by using a large model on selected keyframes and a smaller model on intervening frames.
- Nitro Fusion 2 generates images in one or two diffusion steps and can be run locally with ComfyUI or Diffusers.
Weather, motion, video generation, and image editing4:55
- NVIDIA Strata simulates global atmospheric conditions in 10-minute increments; the model and training recipes are released.
- Flow HMR reconstructs physically plausible 3D human motion from ordinary video. Relift tracks 3D points, including when they are occluded.
- Gemini Omni 1.2 Flash accepts text, image, video, and audio inputs for video generation and editing; GPT Images 2.6 adds GIF creation, markup edits, and SVG export.
Weight-space programming: promise and caveats11:37
- MIT researchers gave a larger model access to smaller models’ weights and, in one experiment, had it change a small fraction of weights to improve behavior.
- A reported math-score gain for Qwen 3.5 was partly tied to fixing repetitive reasoning loops; another analysis found the original model could score much higher with different settings.
- The agent used Python tools and practice questions to test changes, and it did not improve a Gemma model without the same looping issue. The work may help debug models, but does not show universal model upgrading.
3D scenes, open image models, robotics, and voice17:35
- Lucid Lera turns an image into an editable 3D scene; its code was not yet available in the transcript.
- PH3 Image adapts a video model for image generation and editing, including combining up to eight reference images.
- Open Robo Kit bundles libraries for robot planning, motion retargeting, grasping, and agent workflows. Moss Live supports real-time voice conversation with screen or camera input.
Small vision models, robot control, and scientific AI24:08
- Pyramid Jit is an experimental pixel-space image model; the transcript notes it struggles with complex scenes, multiple people, and accurate text.
- Cola 3D creates object-separated 3D meshes from an image, while Mirage is a compact model for identifying objects, reading text, and understanding spatial relationships.
- ZAP controls a robot from camera footage and natural-language instructions after simulation training. Hydrogen 1.0 targets scientific research and is available in versions requiring substantial GPU memory.
New models for security, motion, coding, and multimodal work33:46
- GPT6 Cyber is described as a security-focused model available only through an approved researcher program; Meta Muse adds voice conversations and expressive avatars.
- Mobius generates, describes, completes, and edits human motion; Kimi K3.1 weights were released, but the transcript describes the model as too large for typical consumer hardware.
- Ernie 5.5 Flash, Ripple 65B, and other model releases or previews target reasoning, coding, and multimodal tasks; several were not yet available as weights.
3D environment generation and reconstruction45:35
- VoxLusion generates furnished indoor 3D scenes from a category or simple layout; its code was described as coming soon.
- Solver incrementally reconstructs 3D geometry and camera positions from photos or video, updating the reconstruction as new frames arrive.
- The roundup also mentions Longat 2.5 and Inkling 2 Small for coding, visual understanding, and autonomous tasks, with some releases still pending.
Tools & resources mentioned
- GitHub — Code, installation instructions, and training scripts for several released projects.
- Hugging Face — Model weights and demos mentioned for projects including Hair3D, Strata, PH3 Image, Mirage, Hydrogen, and Kimi K3.1.
- ComfyUI and Diffusers — Local runtimes mentioned for Nitro Fusion 2 and PH3 Image.
- Google AI Studio, API, and Google Flow — Access options named for Gemini Omni 1.2 Flash.
- Tencent Cloud and WorkBuddy — Ways to try Tencent H High4, according to the transcript.
- Tinker — Platform named for trying Inkling 2 Small.
Notes are AI-generated from the video's transcript and may miss nuance.