Hunyuan Image 3.0 tops LMArena for T2V, and it’s fully open-source!

Hunyuan Image 3.0 really takes things to another level, it outperforms both Nano-Banana and Seedream v4, and it’s completely open source!

After testing it myself, I’d say it’s one of the most impressive models I’ve seen for creating artistic or stylized images (aside from Midjourney, of course).

You can dive into the technical breakdown here:
👉 https://github.com/Tencent-Hunyuan/HunyuanImage-3.0

The only real downside at the moment is the size, this thing is enormous. It’s a Mixture of Experts model with around 80B parameters, which makes running it locally a big challenge. That said, the team has an exciting roadmap that includes smaller, distilled versions and new features:

• ✅ Inference
• ✅ HunyuanImage-3.0 Checkpoints
• 🔜 HunyuanImage-3.0-Instruct (reasoning version)
• 🔜 VLLM Integration
• 🔜 Distilled Checkpoints
• 🔜 Image-to-Image Generation
• 🔜 Multi-turn Interaction

Prompt used for the sample image:

“A crystal-clear mountain lake reflects snowcapped peaks and a sky painted pink and orange at dusk. Wildflowers in vibrant colors bloom at the shoreline, creating a scene of serenity and untouched beauty.”
(steps = 28, guidance = 7.5, resolution = 1024×1024)

I also put together a short breakdown video showing results, prompts, and generation examples:
🎥 https://www.youtube.com/watch?v=4gxsRQZKTEs