Extreme AI zoom, new DeepSeek, full body avatars, Kling 2.1, new text-to-speech, robot kickboxing

Key Concepts

Hunyan video avatar, Direct 3DS S2, Gigascale 3D generation, Chain of Zoom, Deepseek R1 0528, Omnic Consistency, Humanoid robot fighting tournament, Phantom by Alibaba, Chat LLM, Deep Agent by Abacus AI, Chatterbox, Paper to Poster, Cling 2.1, EVA (Expressive Virtual Avatars).

Tencent Hunyan Video Avatar

Tencent has released Hunyan video avatar, an AI model that animates a character's image based on audio input, creating realistic lip-syncing, emotions, and movements.

• Functionality: Takes a single image and audio, then animates the character to match the audio.
• Realism: Produces realistic emotions and movements, including head and body motion.
• Multi-Character Support: Can animate multiple characters in a scene with different voices.
• Versatility: Works with realistic photos, drawings, anime, and 3D characters, and even animals.
• Architecture: Uses a multimodal diffusion transformer to combine audio and video.
• Availability: Models are available on Hugging Face.
• Hardware Requirements: Requires an Nvidia CUDA GPU with at least 24 GB of VRAM (96 GB recommended for better quality).
• Open Source: The open-source community is expected to create compressed versions for lower VRAM.

Direct 3DS S2: Gigascale 3D Generation

Direct 3DS S2 is a 3D model generator that creates highly detailed and high-resolution models from a single image.

• Functionality: Generates a 3D model from a 2D image input.
• Detail Level: Produces incredibly detailed and accurate models.
• Comparison: Significantly more detailed than other 3D model generators like Trellis, Hunyen, and High 3D Gen.
• Hugging Face Demo: A free Hugging Face demo is available for users to try it out.
• Efficiency: Uses a spatial sparse attention mechanism, enabling training at 1024 resolution with only eight GPUs.
• Availability: Models and code are available on GitHub.

Chain of Zoom: Extreme Image Upscaling

Chain of Zoom is an AI-powered extreme upscaler that allows for significant image magnification while maintaining sharpness and clarity.

• Magnification: Handles magnification up to 256 times.
• Process: Breaks down the image into smaller parts and uses a vision language model to guide the generation.
• Vision Language Model: Trained using a reinforcement technique called GRPO (generalized reward policy optimization).
• Availability: Code is available on GitHub.
• Hardware Requirements: Can run on a GPU with 24 GB of VRAM (two GPUs recommended).

Deepseek R1 0528 Upgrade

Deepseek has released an upgraded version of its R1 model, Deepseek R1 0528, which shows improved benchmark performance and reduced hallucinations.

• Performance: Beats Google's Gemini 2.5 Pro and OpenAI's flagship 03 model on some benchmark scores.
• Pricing: Cheaper than Gemini 2.5 Pro and GPT-4.
• Open Source: The model is open-source under the MIT license.
• Accessibility: Available for free on their online platform.
• Benchmarks: Performs well on IME (math reasoning) and Live Code Bench (coding).

Omnic Consistency: AI Image Style Transfer

Omnic Consistency is an open-source AI image editor that excels at changing the style of images while preserving details.

• Functionality: Changes the style of a photo while maintaining the original details.
• Comparison: Outperforms other image editors like GBT40, Google's Gemini, and Flux Context.
• Hugging Face Demo: A free Hugging Face demo is available for online testing.
• Availability: Models and datasets are available on GitHub.

Humanoid Robot Fighting Tournament

China held the first-ever humanoid fighting tournament in Hongjo, organized by China Media Group (CMG).

• Robots: Featured four Unitree G1 humanoid robots (1.3m tall, 35kg).
• Control: Robots were teleoperated by humans.
• Autonomous Elements: Robots autonomously regain balance and get back up after falling.

Phantom by Alibaba

Phantom by Alibaba allows users to upload images of characters or objects and insert them into videos, powered by Animate Anyone 2.1.

• Functionality: Inserts images into videos.
• Model Release: The full 14 billion parameter model has been released.
• Integration: A Comfy UI workflow integrates Phantom.

Chat LLM and Deep Agent by Abacus AI

Chat LLM is an all-in-one platform for using various AI models, including image and video generators. Deep Agent is an AI agent capable of performing complex tasks autonomously.

• Chat LLM: Provides seamless switching between different AI models.
• Deep Agent: Can create PowerPoint presentations, find cheap flights, make dinner reservations, and automate workflows.

Chatterbox: Open-Source Text-to-Speech

Chatterbox is a new open-source text-to-speech generator that claims to be better than 11 Labs.

• Functionality: Clones voices from a short audio sample and uses it to speak the given text.
• Voice Cloning: Preserves the tone and expressiveness of the reference voice.
• Availability: The code is available on GitHub.
• Hardware Requirements: Can run on most consumer-grade GPUs, with CPU and Mac support.
• Model Size: Small model with a .5 billion parameter Llama backbone.

Paper to Poster: PDF to Scientific Poster Conversion

Paper to Poster is an AI tool that converts a scientific paper PDF into a conference-ready poster.

• Functionality: Turns a scientific paper into a poster with key findings, visualizations, and figures.
• Comparison: Outperforms other AI models like OpenAI's 40 image generator, GBT40, PPT agent, and OWL.
• Process: Uses a parser, planner, and painter commenter to extract data, layout components, and refine the poster.
• Availability: The code is available on GitHub.
• Cost: Cheaper than other methods, especially when using the open-source Quinn model by Alibaba.

Clling 2.1 Release

Clling has released its latest model, Clling 2.1, with two versions: Master and Normal.

• Cling 2.1 Master: Higher quality but takes longer to generate and costs 100 credits.
• Cling 2.1 Normal: Quality is as good as Clling 2.0, costs 35 credits.
• Quality: On par with the quality of VO3.

EVA (Expressive Virtual Avatars)

EVA creates realistic full-body 3D models of people with accurate body movements, facial expressions, and hand gestures.

• Functionality: Generates 3D models from video input.
• Realism: Produces highly realistic 3D renders.
• Limitations: Movements are based on the input video, and further control is limited.
• Process: Breaks down video into skeleton poses and facial expressions, creates a geometry layer and a 3D Gaussian appearance layer, and refines the face independently.
• Availability: Not yet released.

Conclusion

This week in AI saw significant advancements across various domains, from realistic character animation and detailed 3D modeling to extreme image upscaling and improved language models. Open-source initiatives continue to democratize access to powerful AI tools, while new applications like AI-powered poster generation and humanoid robot competitions showcase the expanding possibilities of AI technology. The advancements in text-to-speech and virtual avatars further highlight the increasing realism and potential of AI in creative and interactive applications.