AI edits videos, full body transfer, insane 3D models, new TTS, Suno v5, new image models - AI NEWS

Key Concepts

Juan Animate, LucyEdit, Kunyan 3D 3.0, SRPO, Yumu, Link Flash 2.0, Tongi Deep Research, Reeve, Wuji hand, Sunno V5, Index TTS2, Vox CPM, Fire Red TTS, Luma Ray 3, Gemini in Google Chrome, AI Marketing Automation Playbook.

Alibaba's Juan Animate: AI Video Generator

• Main Topic: Alibaba released Juan Animate, a powerful open-source video generator based on One 1.2, capable of transferring movements from a reference video onto a new character.
• Key Points:
  • Accurately copies body, face, hand movements, and lip sync.
  • Replaces characters in existing videos while seamlessly blending them with the background.
  • Outperforms similar tools like Animate Anyone and Juan Vase.
  • Superior to Runway Act Two in motion accuracy.
• Examples:
  • Replacing a character in a video with Rick Astley (Rickrolling).
  • Transferring human movements to animated characters and creatures.
• Availability:
  • Available on GitHub with instructions for local download and execution.
  • Requires approximately 40 GB of VRAM.
  • Comfy UI is developing an official workflow.
• Quote: "With this tool, now anyone can act out their own AI movies."

LucyEdit: Free AI Video Editor

• Main Topic: Deart's LucyEdit, a free and open-source AI video editor, allows users to edit videos using text prompts.
• Key Points:
  • Enables microediting of video elements using natural language.
  • Users can change characters, clothing, and appearance.
  • Offers "dev" and "pro" versions.
  • The "dev" version is available as a Comfy UI workflow for offline use.
• Process:
  1. Upload a video to the LucyEdit playground.
  2. Enter a text prompt to edit a specific element (e.g., "change his tit to red").
  3. Generate the edited video.
• Availability:
  • Playground available with free account and 2,000 credits.
  • Comfy UI workflow for Lucydev available on GitHub.
  • The compressed version is 10 GB in size.

Tencent Hunyen 3D: AI 3D Model Generator

• Main Topic: Tencent Hunen released Kunyan 3D 3.0, an advanced AI 3D model generator.
• Key Points:
  • Features three times higher precision and better resolution.
  • Generates ultra HD models with realistic faces and body poses.
  • Accurately predicts and fills in missing parts of the character.
• Process:
  1. Sign up for a free account.
  2. Enter a prompt or upload an image.
  3. Select version 3 and set the model face count.
  4. Generate the 3D model.
• Example: Generating a 3D model from a 2D drawing with high detail and accuracy.
• Availability:
  • Accessible with a free account and 20 free credits.

SRPO: Realistic AI Image Model

• Main Topic: Tencent's SRPO, a fine-tuned version of the Flux AI image model, improves realism and aesthetic quality.
• Key Points:
  • Enhances realism by three times compared to the original Flux model.
  • Generates realistic photos, paintings, and Renaissance art.
  • Outperforms Flux Crea in generating realistic images.
• Examples:
  • Realistic photos of flowers, cats, and people.
  • AI-generated versions of famous paintings like "Starry Night" and "Girl with a Pearl Earring."
• Availability:
  • Available on Hugging Face and GitHub.
  • The model is almost 50 GB in size, but compressed GGUF versions are available.

Yumu: Style and Reference Transfer AI

• Main Topic: Byte Dance's Yumu allows generating photos of reference characters or objects with style transfer.
• Key Points:
  • Transfers the face of a reference character onto a photo.
  • Adds multiple characters to the same photo.
  • Two versions: one based on Uno and the other on Omni Gen 2.
• Process:
  1. Enter a prompt describing the image.
  2. Upload up to three reference images.
  3. Enter a negative prompt.
  4. Specify the dimensions.
• Availability:
  • Free Hugging Face spaces for online use.
  • GitHub repo with instructions for local download and Comfy UI workflows.
  • Model sizes are less than 2 GB.

Link Flash 2.0: Efficient Open-Source Model

• Main Topic: Inclusion AI's Link Flash 2.0, a mixture of experts model, achieves state-of-the-art performance with only 6.1 billion active parameters.
• Key Points:
  • Excels in complex reasoning, code generation, and front-end development.
  • Outperforms larger models in benchmarks like graduate-level science questions and competitive math.
  • Blazing fast, achieving over 200 tokens per second.
• Availability:
  • Available on Hugging Face and Model Scope.
  • Instructions for local download on GitHub.

Tongi Deep Research: AI Research Agent

• Main Topic: Alibaba's Tongi Deep Research, an agentic large language model, matches the performance of OpenAI's deep research.
• Key Points:
  • Autonomously performs multi-step research, including web searching and code execution.
  • Outperforms closed proprietary models like Gemini Deep Research and OpenAI deep research.
  • Uses only 3 billion active parameters.
  • Offers a "deep research heavy mode" for complex tasks.
• Examples:
  • Identifying a university based on complex criteria.
  • Solving PhD-level math questions.
• Availability:
  • Model available on Hugging Face (around 60 GB).
  • Instructions for local download on GitHub.

Reeve: AI Image Editor

• Main Topic: Reeve, an AI image editor, allows users to edit images seamlessly with object detection and microediting capabilities.
• Key Points:
  • Automatically detects objects and their positions in the image.
  • Allows users to adjust the size and proportions of each object.
  • Enables microediting of specific parts of the image without affecting the rest.
• Process:
  1. Upload an image.
  2. Press "edit" to automatically detect objects.
  3. Adjust the size and proportions of objects or edit specific parts.
• Availability:
  • Accessible with a free account.

Other AI Tools and Updates

• Wuji Hand: A robotic hand by Wuji Tech with 20 active degrees of freedom and tactile sensors.
• Sunno V5: A preview of the upcoming version 5 of the AI music generator Sunno, promising improved vocals.
• Index TTS2: A text-to-speech generator good at expressive generations with emotion control.
• Vox CPM: A text-to-speech generator that clones voices with a few seconds of reference audio and detects emotions and accents.
• Fire Red TTS: A text-to-speech generator that supports up to four speakers, multilingual support, and long generations.
• Luma Ray 3: Luma Labs' latest video model, Ray 3, capable of thinking and reasoning, outputting 16-bit high dynamic range color.
• Gemini in Google Chrome: Integration of Gemini in Google Chrome, allowing users to access Gemini directly from the browser and refer to the current tab.

HubSpot's AI Marketing Automation Playbook

• Main Topic: A free guide from HubSpot Media and Masters in Marketing to supercharge marketing with AI.
• Key Points:
  • Streamlines workflows, boosts engagement, and drives better results with AI.
  • Provides actionable steps to integrate AI into marketing strategy.
  • Includes a workflow audit, a guide to choosing the right AI tools, and strategies for automating and personalizing content.
  • Leverages AI for smarter analytics and converting leads into revenue.

Conclusion

This week in AI has seen significant advancements across various domains, including video generation, image editing, 3D modeling, text-to-speech, and research agents. Open-source models like Juan Animate, LucyEdit, SRPO, Link Flash 2.0, and Tongi Deep Research are pushing the boundaries of what's possible, often rivaling or surpassing proprietary models. Tools like Kunyan 3D 3.0, Yumu, Reeve, Vox CPM, and Fire Red TTS offer new creative possibilities for users. Google's integration of Gemini into Chrome and HubSpot's AI Marketing Automation Playbook highlight the increasing accessibility and practical applications of AI in everyday life and business.