Text-to-Video generation with Hunyuan Video 1.5 in ComfyUI

Step-by-step

1. Load the Hunyuan Video 1.5 main model, selecting FP16 for quality or FP8 if you have low VRAM.
2. Enable Sage Attention to increase generation speed and quality.
3. Load the dual text encoders, specifically selecting the Qwen2.5-VL and BYT5 models.
4. Set the model type to 'hunyuan VIDEO 15' in the encoder node.
5. Configure the latent node resolution to 1280x720 with a frame length of 33.
6. Set the CFG to 6.0 and the Shift value to 9 for 720p generation.
7. Configure the KSampler with 20 steps, the Euler sampler, and the Simple scheduler.

Options

• Use FP8 instead of FP16 to avoid out-of-memory errors.
• Use 50 sampling steps for higher quality if time is not a concern.
• Write prompts manually or use the Qwen VL model to generate them from an image.

Watch out for

• You must select 'hunyuan VIDEO 15' in the text encoder type or it will not work correctly.
• The Shift value must be adjusted according to the specific resolution list (e.g., 9 for 720p).

Tips

• Enable Sage Attention for a faster and higher-quality result.
• Use 20 sampling steps instead of 50 to save significant time while maintaining good results.
• Refer to the official GitHub guide for the specific prompt engineering format.
• Use a prompt engineering workflow with Qwen VL for better descriptive prompts.

Highlights

“we can then achieve a fantastic effect”