DeepSeek is back, realtime upscaler, realtime 3D worlds, top 3D generator, Qwen Image 2512: AI NEWS

AI Weekly Roundup: Breakthroughs in Video, 3D, and Image Generation (Recent Developments)

Key Concepts:

• Stream Diff VSSR: Real-time video upscaling AI.
• UltraShape 1.0: Open-source 3D model generator from images.
• Ume 1.5: AI for creating explorable, interactive 3D worlds.
• MHC (Manifold Constraint Hyperconnections): New architecture for DeepSeek models, improving performance and efficiency.
• Spot Edit: AI for selective image editing, modifying specific areas without full regeneration.
• HY-Motion 1.0 O: Tencent AI for generating 3D animations from text prompts.
• TwinFlow: Acceleration method for Alibaba’s Zimage Turbo, significantly reducing generation time.
• ProEedit: AI capable of editing both images and videos with text prompts.
• Javis/Javis GPT: Multimodal AI that processes and generates video, audio, and text.
• Space-Time Pilot: AI for re-shooting videos with altered perspectives and movements.

1. Real-Time Video Upscaling with Stream Diff VSSR

A new AI, Stream Diff VSSR, offers real-time video upscaling capabilities. Demonstrations show significant improvements in video quality, sharpness, and detail, even with shaky footage. Compared to traditional diffusion VSSR, which takes approximately 4600 seconds for upscaling, Stream Diff VSSR achieves results in roughly 0.3 seconds on an RTX 4090 (24GB VRAM). The code is publicly available on GitHub, with plans to release the training code as well. ([GitHub Repo Link in Description])

2. High-Fidelity 3D Model Generation with UltraShape 1.0

UltraShape 1.0 is a new open-source AI capable of generating highly detailed 3D models from 2D images. Examples demonstrate its ability to accurately render complex objects like dragons and human characters. While it currently generates only the shape of the model (not texture or color), it outperforms other open-source competitors like Trellis 2 in terms of detail and accuracy. The code and data preparation instructions are available on GitHub. ([GitHub Repo Link in Description])

3. Interactive 3D World Creation with Ume 1.5

Ume 1.5 allows users to create explorable, interactive 3D worlds from images or text prompts. These worlds can be navigated in real-time using standard controls (WASD/arrow keys). The AI utilizes a masked video diffusion transformer for efficient video generation and supports acceleration methods like bidirectional attention distillation and self-forcing for real-time performance. It can also inject events or characters based on text prompts. A one-click installer for Windows is available, and the model runs effectively on a laptop GPU with 16GB of VRAM. The code for data preparation and model training is also released. ([GitHub Repo Link in Description])

4. DeepSeek’s MHC: Improving Model Performance and Stability

DeepSeek has released a new paper detailing MHC (Manifold Constraint Hyperconnections), a novel architecture designed to improve the performance and stability of deep transformer models. Traditional transformers use residual connections (shortcuts) to prevent signal loss in deep layers. Hyperconnections expand this concept with multiple lanes for data flow, but can lead to instability. MHC addresses this by implementing a “manifold constraint,” essentially an energy conservation law that regulates information flow and prevents chaotic signal behavior. This, combined with a recomputing trick to reduce memory usage, results in a more performant and efficient model, outperforming standard transformers and models with basic hyperconnections on benchmark tests. ([Paper Link in Description])

5. Selective Image Editing with Spot Edit

Spot Edit enables targeted image editing, allowing users to modify specific areas of an image without regenerating the entire image. This results in faster and more consistent edits. For example, changing a soccer ball to a sunflower without affecting other elements. This contrasts with tools like Nano Banana, which typically regenerate the entire image. The architecture consists of a "spot selector" and a "spot fusion" component. Code is available on GitHub, with support for Flux Context and Quen imageedit. ([GitHub Repo Link in Description])

6. Tencent’s HY-Motion 1.0 O: Realistic 3D Animation Generation

Tencent’s HY-Motion 1.0 O is a 1 billion parameter text-to-motion model capable of generating realistic 3D animations of characters performing various actions. The model was trained on over 3,000 hours of motion data, refined with 400 hours of high-quality data, and further improved with reinforcement learning using human feedback. It can also animate characters interacting with props. The model is available on GitHub, requiring at least 26GB VRAM for the standard model and 24GB for a lightweight version. ([GitHub Repo Link in Description])

7. Accelerated Image Generation with TwinFlow

TwinFlow is an acceleration method for Alibaba’s Zimage Turbo, reducing image generation time by up to 5x. It achieves this by reducing the number of steps required for image generation from 7-9 to 1-2. While currently lacking Comfy UI integration, it offers significant speed improvements. ([Hugging Face Repo Link in Description])

8. Unified Image and Video Editing with ProEedit

ProEedit is a multimodal AI capable of editing both images and videos using text prompts. It can remove objects, add elements, and modify scenes in both media types. While video editing results may exhibit increased saturation, the ability to apply the same model to both images and videos is a significant advancement. Code for both image and video editing is planned for release. ([Project Page Link in Description])

9. Miniax M2.1: Top Open-Source AI Model

Miniax M2.1, sponsored by Miniax, is a new open-source AI model rivaling top closed models like Gemini 3 Pro and GPT-5.2 in coding, tool use, and multilingual tasks. It excels in agentic coding and outperforms competitors on benchmarks like SWEBench. The model is available for local download and online use with an exclusive discount for viewers. ([Link in Description])

10. HighStream: 100x Faster Video Generation

HighStream, developed by Alibaba, accelerates video generation using Alibaba 1 by up to 107x. It achieves this by processing video in chunks, utilizing an anchor-guided sliding window, and employing spatial compression. While the code is currently under legal review, the technology demonstrates significant potential for faster AI video creation. ([Technical Paper Link in Description])

11. IQ Quest Coder V1: Powerful Coding AI

IQ Quest Coder V1, backed by a Chinese quant trading firm, is a powerful coding AI available in 17-40 billion parameter versions with a 128,000 token context window. Trained using a code flow method, it excels in agentic tasks and multi-step reasoning, achieving competitive results on benchmarks like SWEBench. The models are open-sourced and available on Hugging Face. ([GitHub Repo Link in Description])

12. Quen Image 2512: Improved Image Generation

Alibaba’s Quen Image 2512 is a new image generator that produces more realistic images with improved text rendering and prompt following compared to its predecessor. It outperforms Zimage Turbo in several tests and is integrated into Comfy UI. Quantized versions are available for lower VRAM GPUs. ([Hugging Face Repo Link in Description])

13. Javis/Javis GPT: Multimodal AI for Video, Audio, and Text

Javis/Javis GPT is a multimodal AI capable of analyzing and generating video, audio, and text. It can understand video content, answer questions about it, and generate new videos with audio based on text prompts. The code is available on GitHub. ([GitHub Repo Link in Description])

14. Space-Time Pilot: Reshooting Videos with Altered Perspectives

Space-Time Pilot allows users to reshoot existing videos with altered perspectives, movements, and effects like bullet time. It can change camera angles, slow down footage, and create dynamic visual effects. The code and datasets are currently under internal review for release. ([Technical Paper Link in Description])

Conclusion:

This week showcased significant advancements across multiple areas of AI, particularly in video and 3D generation. The release of open-source models like UltraShape 1.0, Ume 1.5, and IQ Quest Coder V1, alongside improvements to existing tools like Zimage Turbo and Alibaba 1, demonstrates a rapidly evolving landscape. The focus on efficiency, stability, and multimodal capabilities signals a trend towards more powerful and accessible AI tools for creators and developers.