ComfyUI Hunyuan Video Low VRam with Upscaling Workflow

Today, I’ll walk you through my experience using HunyuanVideo with Comfy UI and how I improved my results. Let’s get started!

So, I tried generating a video in ComfyUI, but the results weren’t great. The video looked blurry and lacked the sharpness I was hoping for. Meanwhile, I had another video processing on the HunyuanVideo website, and it already looked so much better. That’s when I realized: maybe my settings needed some tweaking.

Upscaling is like giving your video a makeover. It enhances the resolution and makes everything look crisp and professional. Here’s how I did it:

1. Load the Video:
   I selected the video I had just generated in Comfy UI.
2. Set the Upscale Factor:
   I chose an upscale factor of 2, which doubles the resolution.
3. Choose the Model:
   I used the 4x Pool Hard Model for upscaling. This model is great for adding clarity and sharpness to videos.
4. Process the Video:
   After running the upscale process, the video’s quality improved noticeably. The colors popped, and the details were even more refined.

I decided to experiment with the settings in Comfy UI, and here’s what worked for me:

1. Resolution:
   I set the resolution to 512 by 512 pixels. This size is manageable and still produces decent-quality videos.
2. Frame Steps:
   I adjusted the steps to 30. This controls how detailed the video gets during processing.
3. Embedding Guidance Scale:
   Setting this to 6 helped align the video more closely with my text prompt.
4. BF16 Mode:
   I made sure all settings used BF16, which is a precision mode for faster and more efficient processing.
5. Video Settings:
   • Bitrate: I kept this at 64 kbps.
   • Frame Rate: Adjusted to 16 fps for smooth playback.

Sometimes, things don’t go as planned. If you encounter an error like “device allocation failed,” don’t panic.
Here’s a quick fix:

• Cancel the process.
• Wait for 5 to 10 seconds.
• Try generating the video again.

To make sure everything runs smoothly, you’ll need to download a few files and place them in the right folders in Comfy UI. Don’t worry—I’ll guide you step by step!

1. Transformer and VAE Files

• Download Location: Transformer and VAE Files on Hugging Face
• Where to Place Them:
  • Save the hunyuan_video_720_cfgdistill_fp8_e4m file in the UNET folder in your ComfyUI directory.
  • Save the hunyuan_video_vae_bf16.safetensors file in the vae folder in your Comfy UI directory.

These files won’t autodownload, so you need to grab them manually from the link above.

2. LLM Text Encoder

• Download Location: LLM Text Encoder on Hugging Face
• Where to Place It:
  • Go to ComfyUI/models/LLM/ open in cmd

git clone https://huggingface.co/Kijai/llava-llama-3-8b-text-encoder-tokenizer

3. Clip Text Encoder

• Download Location: Clip Text Encoder on Hugging Face
• Open in cmd ComfyUI/models/
• type this and enter

git clone https://huggingface.co/openai/clip-vit-large-patch14

This also supports auto download, but you can manually download the files if needed.

Here’s a quick tip:

The good news? This model can create functional videos even at low resolutions, so you don’t need a supercomputer to get started!

The memory usage depends on the resolution and frame count of your video.

Even on a system with 12 GB of VRAM, you might hit some limits if you push the resolution too high.