HunyuanVideo: High-Quality Long-Form Generation

HunyuanVideo: Tencent's Quality Leader

HunyuanVideo from Tencent produces some of the highest-quality video from a local model in 2026. While Wan 2.2 is the most commonly used model due to its native ComfyUI support and accessible VRAM requirements, HunyuanVideo's output quality — particularly for smooth motion, coherent physics, and cinematic aesthetics — is consistently rated at or above the top proprietary video models.

The original HunyuanVideo model required 60 GB+ VRAM in its native form. HunyuanVideo 1.5, released in early 2026, reduced the parameter count to 8.3 billion while retaining most of the quality advantages. HunyuanVideo 1.5 runs on consumer GPUs starting at 8 GB VRAM with aggressive quantization, and produces good results at 16 GB+ with moderate quantization.

Installation Options

HunyuanVideo has three installation paths in ComfyUI:

• Native ComfyUI support: if you have a recent ComfyUI version, drag the official HunyuanVideo workflow JSON from docs.comfy.org/tutorials/video/hunyuan/hunyuan-video onto the canvas. ComfyUI will list missing models and prompt you to download them via the Missing Nodes panel.
• Kijai's HunyuanVideoWrapper: install via Manager (search "HunyuanVideoWrapper"). This provides additional nodes with more configuration options and active updates. Recommended if you plan to use HunyuanVideo regularly.
• GGUF quantized via ComfyUI-GGUF: install ComfyUI-GGUF from Manager, then download HunyuanVideo 1.5 GGUF models from Hugging Face (search "city96/HunyuanVideo-GGUF"). Q8_0 gives near-full quality at lower VRAM; Q4_K_M halves VRAM needs with noticeable but acceptable quality loss.

Model Files and Placement

HunyuanVideo requires multiple model files:

• Main model: hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors (or GGUF equivalent) → ComfyUI/models/diffusion_models/
• VAE: hunyuan_video_vae_bf16.safetensors → ComfyUI/models/vae/
• LLM text encoder (LLaVA / mT5): used for text conditioning → ComfyUI/models/LLM/ or ComfyUI/models/clip/ depending on the wrapper version you installed
• CLIP-L: clip_l.safetensors → ComfyUI/models/clip/

The easiest path: load the official ComfyUI workflow template, click "Install Missing Models," and let the download manager fetch everything to the correct locations automatically.

Key Workflow Parameters

The HunyuanVideo workflow (via Kijai's wrapper) centers on the HunyuanVideoSampler node. Key parameters:

• width / height: 720×1280 (portrait) or 1280×720 (landscape) for standard outputs. 544×960 for lower VRAM.
• num_frames: 53 frames at 24 FPS = approximately 2.2 seconds. 125 frames ≈ 5 seconds. Increase carefully — VRAM scales with frame count.
• steps: 30 steps default; 20 steps for faster iteration at slightly lower quality
• cfg: 1.0 (HunyuanVideo uses embedded guidance; classifier-free guidance is not applied the same way as in SD models)
• embedded_guidance_scale: 6.0 default; controls how strongly the model follows the text prompt. Increase to 7–8 for more literal prompt following.

Prompting HunyuanVideo

HunyuanVideo uses an LLM-based text encoder that understands natural language at a high level — similar to FLUX. Write detailed, descriptive prompts in paragraph form rather than tag lists:

"A slow aerial drone shot descends toward a fog-covered mountain valley at dawn, pine trees emerging from white mist, warm golden light filtering through the clouds, muted greens and blues, photorealistic, cinematic aspect ratio."

Include camera movement descriptions (dolly in, slow pan right, aerial descent) as these significantly affect the generated motion. HunyuanVideo responds to camera language more reliably than most other video models.