🪄 Introducing daVinci-MagiHuman: The Performance-Level Audio-Video Generative Foundation Model Proudly open-sourced and jointly developed by SII GAIR Lab & Sand.ai, it sets a new standard for multimodal AI. ⏳ 1/6
daVinci MagiHuman Text / Image to Video Generator with Audio Sync
Create videos with daVinci MagiHuman - a 15B open-source audio-video foundation model by Sand.ai and SII GAIR Lab. Generate synchronized video and audio from text or images with industry-leading lip sync accuracy across 7 languages. Supports up to 1080p resolution with 5-10 second duration. Powered by a single-stream Transformer architecture with no cross-attention, delivering 5-second 256p video in just 2 seconds on a single H100.
Public
daVinci MagiHuman Text to Video Gallery
Experience the cinematic power of daVinci MagiHuman text-to-video generation. Create stunning videos with synchronized audio from detailed text descriptions, featuring industry-leading lip sync across 7 languages.
AI Video
Rainy Tokyo Night
A woman in a red coat walks through a neon-lit Tokyo alley on a rainy night with shimmering reflections.
Prompt
“Rainy night in a neon-lit Tokyo alley, a woman in a red coat walks slowly under an umbrella. Reflections shimmer on wet cobblestones. Handheld camera follows her from behind, bokeh street lights, cinematic color grade, moody atmosphere.”
daVinci MagiHuman Image to Video Gallery
Transform your static images into dynamic videos with daVinci MagiHuman. Experience seamless image-to-video conversion with realistic facial expressions, natural body motion, and synchronized lip-synced audio.
Input
Output
Podcast Host Speaking
daVinci MagiHuman YouTube Videos
Watch community demonstrations and reviews showcasing daVinci MagiHuman's audio-video generation capabilities
- daVinci-MagiHuman: Fast Audio-Video Synthesis - AI Research Roundup
- 达芬奇最新开源模型，革命Seedance2.0 DaVinci-MagiHuman：开源音视频生成新标杆，5秒视频2秒出，还能说6种语言！ - XIAOXIAO LI
- LTX 2.3, Veo и Sora больше не нужны? Тестируем daVinci-MagiHuman - ServerFlow AI Lab - R&D в области ИИ и LLM
- Ai动画224-化繁为简！daVinci-MagiHuman，快速音视频生成基础模型的单流架构，支持多国语言，音画同步，音色参考-T8 Comfyui教程 - T8star-Aix
- New OpenSource Video Model, #1 Image generator, Seedance 2.0 Drop, replit and lovable in danger - AI Research
daVinci MagiHuman YouTube Videos
Watch community demonstrations and reviews showcasing daVinci MagiHuman's audio-video generation capabilities
daVinci MagiHuman Popular Reviews on X
See what people are saying about daVinci MagiHuman on X (Twitter)
daVinci-MagiHuman is a 15B single-stream Transformer, trained from scratch to generate synced video+audio with self-attention only—no cross-attention or multi-stream paths. It is open-source, supports 6 languages, beats Ovi/LTX, and runs on one H100.
I have been testing open source daVinci-MagiHuman, a single-stream 15B Transformer trained from scratch that jointly generates video + audio. 5s 1080p video in 38s on a single H100, about 1 minute on newer gaming Nvidia GPUs By @SII_GAIR + @SandAI_HQ
daVinci-MagiHumanという新しい動画生成モデルがオープンで出た。これがLTX-2.3よりもすごいとかいう話。特にオーディオ生成がいい感じらしい。さらに多言語対応してて日本語の音声も対応してると書かれてる。開発したGAIRってのは上海イノベーション研究所内の研究ラボらしい reddit.com/r/StableDiffus…
映像と音声を同時生成のオープンソースモデル「daVinci-MagiHuman」が登場 ・OSS界隈ではトップクラスの性能 ・日中英韓独仏の6言語対応 ・音声認識誤り率14.6% クローズドのSeedance 2.0に対抗。デモの感じは精度が高そう H100で5秒間の1080p動画を38秒で生成したらしい
動画生成AIはオープンソースでも戦えるか？ daVinci-MagiHuman は、動画と音声をシングルストリームの15B Transformerで同時生成する完全オープンソースモデル。 Ovi 1.1に80.0%、LTX 2.3に60.9%勝率。 H100で1080pの5秒の動画を38.4秒で生成。日本語にも対応！ 詳細は🧵
DaVinci-MagiHuman for ComfyUI. - 15B-param single-stream model runs in ~6GB VRAM via block-level swapping; - 8-step distillation; github.com/mjansrud/Comfy…
daVinci-MagiHuman. We have another fast single-stream audio-video 15B foundation model by @SandAI_HQ > no separate pathways or cross-attention modules. > just raw self-attention doing all the heavy lifting. > wins 80% vs Ovi 1.1, 60% vs LTX 2.3; > native multilingual realistic
What's daVinci MagiHuman
Sand.ai's 15B open-source audio-video foundation model with best-in-class lip sync
15BParameters
1080pMax Resolution
7Languages Supported
2s256p Generation Speed
daVinci MagiHuman is a 15-billion parameter single-stream Transformer that jointly generates synchronized video and audio from text or images, achieving industry-leading lip sync accuracy with a 14.6% word error rate across 7 languages.
15BParameters
1080pMax Resolution
7Languages Supported
2s256p Generation Speed
What's daVinci MagiHuman
Sand.ai's 15B open-source audio-video foundation model with best-in-class lip sync
daVinci MagiHuman is a 15-billion parameter single-stream Transformer that jointly generates synchronized video and audio from text or images, achieving industry-leading lip sync accuracy with a 14.6% word error rate across 7 languages.
daVinci MagiHuman's Powerful Features
Discover the advanced capabilities that make daVinci MagiHuman exceptional for audio-video generation
Joint Audio-Video Generation
Generate synchronized video and audio in a single pass using a unified single-stream Transformer architecture with self-attention only, eliminating the need for separate audio pipelines.
Industry-Leading Lip Sync
Achieve 14.6% word error rate for lip synchronization, significantly outperforming competitors like Ovi 1.1 (40.45%) and LTX 2.3 (19.23%) in speech accuracy benchmarks.
7-Language Speech Support
Generate speech-synchronized videos in English, Chinese (Mandarin and Cantonese), Japanese, Korean, German, and French with natural pronunciation and lip movements.
Ultra-Fast Generation
Produce a 5-second 256p video in just 2 seconds on a single H100 GPU. 8-step DMD-2 distillation eliminates the need for classifier-free guidance without quality loss.
Dual Input Modes
Create videos from text prompts or animate still images. Both text-to-video and image-to-video modes support configurable aspect ratios, resolutions, and durations from 5 to 10 seconds.
Up to 1080p Super-Resolution
Generate videos at 256p, 540p, 720p, or 1080p through a latent-space super-resolution pipeline that upscales without extra VAE decode-encode overhead for efficient high-res output.
Open Source Apache 2.0
Fully open-sourced under Apache 2.0 license with complete stack including base weights, distilled model, super-resolution model, and inference code for unrestricted commercial use.
Human-Centric Excellence
Specialized in digital human generation with expressive facial performance, realistic body motion, and consistent character preservation across frames for professional talking-head content.
Frequently Asked Questions
Common questions about daVinci MagiHuman audio-video generation
Still have questions?
daVinci MagiHuman supports two primary input modes: Text-to-Video (generate videos with synchronized audio from text prompts) and Image-to-Video (animate still images into motion videos with optional audio). Both modes support configurable aspect ratios (16:9 landscape, 9:16 portrait), resolutions up to 1080p, and durations from 5 to 10 seconds.
daVinci MagiHuman supports synchronized speech generation in 7 languages: English, Chinese (Mandarin), Cantonese, Japanese, Korean, German, and French. The model achieves a 14.6% word error rate for lip synchronization, significantly outperforming competitors like Ovi 1.1 (40.45% WER) and LTX 2.3 (19.23% WER).
daVinci MagiHuman supports multiple resolutions: 256p (fastest), 540p (super-resolution), 720p, and 1080p (super-resolution). Video duration can be configured from 5 to 10 seconds with 1-second granularity. Both landscape (16:9) and portrait (9:16) aspect ratios are supported.
On a single NVIDIA H100 GPU, daVinci MagiHuman generates a 5-second 256p video in approximately 2 seconds. For higher resolutions, generation times increase: 540p takes about 8 seconds and 1080p takes about 38.4 seconds for a 5-second video. This speed is achieved through 8-step DMD-2 distillation that eliminates classifier-free guidance.
Yes, daVinci MagiHuman is fully open-sourced under the Apache 2.0 license by Sand.ai and SII GAIR Lab. The complete stack is available including base model weights, distilled model, super-resolution model, and inference code, allowing unrestricted commercial use, modification, and distribution.
daVinci MagiHuman stands out with its unique single-stream Transformer architecture that uses self-attention only (no cross-attention or multi-stream paths), enabling joint audio-video generation in a single model. It achieves best-in-class lip sync accuracy (14.6% WER), supports 7 languages for speech, and an 80% win rate against Ovi 1.1 in human evaluation of visual quality.
How to Use daVinci MagiHuman Text to Video
Generate videos with synchronized audio from text descriptions
1
Write Your Prompt2
Configure Settings3
Generate VideoEnter a detailed description of the video you want to create. Include subject, action, speech content, and desired language for best lip-synced results.
How to Use daVinci MagiHuman Image to Video
Animate still images into videos with synchronized audio
1
Upload Your Image2
Add Prompt and Settings3
Generate Animated VideoUpload a reference image of the person or scene you want to animate. daVinci MagiHuman excels at human-centric content with realistic facial expressions and body motion.
Flexible AI Pricing
Pay-as-you-go credits or subscription plans. No hidden fees, cancel anytime.
Monthly billing