High-Res Image-to-Video with LoRA
HunyuanVideo-I2V, from Tencent, is a open-source image-to-video generation model. Up to 720p resolution, 129 frames. Supports custom LoRA training for unique effects.
The product includes LoRA (Low-Rank Adaptation) training code that allows users to fine-tune the model for specific visual effects (e.g., hair growth, embrace effects) while keeping base model weights frozen, reducing computational overhead and training time.
HunyuanVideo-I2V converts a reference image into video by processing the image through a multimodal language model to extract semantic tokens, then combines these with video generation to maintain first-frame consistency while animating the scene according to text prompts. Supports up to 720p resolution and 129 frames (5 seconds).
HunyuanVideo-I2V provides PyTorch model definitions, pre-trained weights, and inference code that can be imported and called programmatically via Python. Supports both single-GPU inference and parallel multi-GPU inference via xDiT for scalability.
The model uses token replacement technique to reconstruct reference image information into the generation process, with stability modes (--i2v-stability flag) and flow-shift parameters to control how closely the first frame is preserved versus allowing dynamic motion.
Product supports CPU offloading, parallel inference across multiple GPUs via xDiT, and integrates with community tools like ComfyUI for non-technical deployment. Includes Docker images pre-configured with CUDA 12.4 for reproducible environments.