A ComfyUI custom node implementation for optimized video generation and motion effects, designed to work with Hunyuan text-to-video models.
Image to Video:
Text to Video:
An optimized video sampler that extends ComfyUI's KSampler capabilities:
- Memory-efficient batch processing for video frames
- Progress tracking for long video generation tasks
- Optimized VRAM usage through sequential frame processing
- Interrupt-safe with proper memory management
- Compatible with all standard ComfyUI samplers and schedulers
A powerful tool for creating motion sequences from static images:
- Horizontal panning effects with adjustable range
- Progressive zoom capabilities
- Seamless loop generation through mirror techniques
- Configurable frame count and motion parameters
A specialized resizing tool optimized for Hunyuan video generation:
- Predefined aspect ratios optimized for home GPUs
- Multiple size options for each aspect ratio
- All dimensions properly aligned to 16x16 grid
- Multiple upscaling methods
- Crop control options
A latent initialization tool specifically designed for Hunyuan video generation:
- Supports multiple optimized resolutions for home GPUs
- Common aspect ratios (16:9, 4:3, 3:2, 9:16, 3:4, 2:3)
- Memory-efficient latent generation
- Configurable video length and batch size
- All dimensions automatically aligned to model requirements
- Clone this repository into your ComfyUI custom nodes directory:
cd ComfyUI/custom_nodes/
git clone https://github.com/ShmuelRonen/ComfyUI-HunyuanVideoSamplerSave.git- Restart ComfyUI to load the new nodes.
-
Image Motion Setup
- Input: Any source image
- Configure motion parameters:
- move_range_x: Controls horizontal movement (-150 to 150)
- frame_num: Number of frames to generate (2 to 500)
- zoom: Progressive zoom effect (0.0 to 0.5)
- Output: Sequence of motion-affected images
-
Image Resizing
- Use ResizeImageForHunyuan to ensure proper dimensions
- Select from optimized presets for your GPU
- Choose appropriate upscaling method
-
Latent Setup
- Use EmptyVideoLatentForHunyuan to initialize latent space
- Select resolution from optimized presets
- Configure video length and batch size
-
Video Generation
- Use HunyuanVideoSamplerSave with your text prompts
- The motion-influenced latents guide the video generation
- Adjustable parameters:
- Steps: Generation steps per frame
- CFG: Prompt influence strength
- Sampler and Scheduler selection
- Denoising strength
- model: Loaded Hunyuan model
- positive/negative: Conditioning from text prompts
- video_latents: Input latent sequence
- seed: Generation seed for reproducibility
- steps: Number of sampling steps
- cfg: Conditioning strength
- sampler_name: Choice of sampling algorithm
- scheduler: Noise scheduler selection
- denoise: Denoising strength
- image: Input source image
- move_range_x: Horizontal motion range
- frame_num: Number of frames to generate
- zoom: Zoom effect intensity
- image: Input image to resize
- size_preset: Selection of predefined sizes (e.g., "384x216 (16:9)", "768x432 (16:9)")
- upscale_method: Choice of upscaling algorithm (nearest-exact, bilinear, area, bicubic)
- crop: Crop method selection (disabled, center)
- resolution: Selection of optimized video resolutions
- length: Number of frames to generate
- batch_size: Number of videos to generate in parallel
The nodes implement several memory optimization strategies:
- Sequential frame processing
- Active memory management
- Intermediate result storage
- Garbage collection during processing
- Optimized resolution presets for home GPUs
- Proper dimension alignment for efficient processing
This allows for processing of longer sequences and higher resolution outputs compared to standard sampling approaches.
This custom node is designed to work seamlessly with:
- ComfyUI's core components
- Hunyuan text-to-video models
- Standard VAE encoders
- Various sampling and scheduling methods
- ComfyUI installation
- Compatible Hunyuan model
- Sufficient VRAM for video processing
- Python 3.x
- PyTorch
MIT
