March 6, 2025 – Tencent Open Source Lab released Hunyuan Video v1.2 featuring groundbreaking Image-to-Video (I2V) generation capabilities with native ComfyUI workflow support. The update garnered 2.4k GitHub stars within 12 hours of release.
Core Technical Specifications
Architectural Innovation
- 13B Hybrid Architecture: Built on Video Diffusion Transformer framework
- Dual-Modal Control: Supports CLIP image embeddings + text prompts (3:2 ratio)
- Memory Optimization: 40% VRAM reduction via gradient checkpointing
System Requirements
| Specification | Minimum | Recommended |
|---|---|---|
| GPU | RTX 3090 | A100 40GB |
| RAM | 64GB | 128GB |
| Output Length | 3s@768×512 | 5s@512×512 |
| Render Time | ~90s/frame | ~30s/frame |
Optimized Workflow (ComfyUI v3.1.6+)
Node configuration
hunyuan_loader = HunyuanVideoLoader(model_path="hunyuan_i2v_v1.2.safetensors")
image_preproc = ImageNormalizeNode(target_size=(512,768), keep_ratio=True)
video_output = VAEDecodeNode(fps=24, format="mp4")Key Innovations
- Temporal Consistency : 62% motion jitter reduction via optical flow guidance
- LoRA Training : 3 modes (Style/Resolution/Motion control)
- Multi-Stage Rendering : Coarse-to-Fine strategy improves 4x efficiency
Troubleshooting Guide
Q: Black screen in output?
- Verify RGB input format
- Adjust CLIP skip (2-3 recommended)
- Try 512×512 resolution
Q: Speed optimization tips?
- Enable xFormers
- Use –medvram flag
- Switch to FP16 precision
Resources :
- [Documentation]： https://comfy.org/hunyuan-docs-en
- [Showcase]： https://comfy.org/hunyuan-showcase-en
- [GitHub]： https://github.com/Tencent/HunyuanVideo
