Tencent published the HunyuanVideo technical report and open-sourced its 13B parameter weights, pairing a 3D causal VAE and dual-to-single-stream DiT with a Multimodal LLM prompt encoder to deliver frontier motion quality to local ComfyUI and LoRA workflows.
HunyuanVideo
Also known as: Tencent HunyuanVideo · HunyuanVideo-I2V · HunyuanVideo-Avatar
Plain English
Tencent's open-weight 13-billion-parameter video generation model family that helped close the quality gap between closed commercial video APIs and locally runnable open checkpoints.
Technical Overview
HunyuanVideo pairs a 3D causal VAE (4x8x8 spatiotemporal compression) with a unified dual-stream-to-single-stream DiT trained under flow matching (Diffusion Models) and conditioned on a decoder-only Multimodal LLM text encoder rather than CLIP/T5 alone.
Timeline (4)
2024
Release2025
ReleaseTencent released HunyuanVideo-I2V, extending the 13B foundation backbone to image-to-video synthesis up to 720p alongside official LoRA fine-tuning code for custom motion and visual effects.
- Paper
Tencent introduced HunyuanCustom, a subject-consistent video customization framework built on HunyuanVideo that fuses text, reference images, audio, and video conditions via a LLaVA-based multimodal understanding module and identity-injection blocks.
- Paper
Tencent published HunyuanVideo-Avatar, adapting the DiT backbone for audio-driven multi-character human animation with dynamic facial emotion and full-body motion preservation from a single portrait reference.
See also (14)
Open full graphHunyuanVideo
9 more connections
- Latent Space & VAEOutgoing · Uses
- Video Generation Models (Sora, Veo, Kling, Wan)Outgoing · Part ofIncoming · Mentions
- CogVideo & CogVideoXIncoming · Alternative to
- CheckpointsOutgoing · Mentions
- Chinese AI LabsIncoming · Mentions
- Diffusion ModelsOutgoing · Mentions
- Hugging Face DiffusersIncoming · Mentions
- LoRA & QLoRAOutgoing · Mentions
- Multimodal & Vision-Language ModelsOutgoing · Mentions