HunyuanVideo 1.5

Where Tencent's original HunyuanVideo topped 13 billion parameters, version 1.5 shrinks to 8.3 billion: a Diffusion Transformer coupled to a causal 3D VAE that compresses frames 16x spatially and 4x in time, so generation fits on one consumer GPU. A single codebase generates clips from a written prompt or from a starting image at 480p or 720p, defaulting to 121 frames, about five seconds at 24 fps, and a separate distilled upscaling model can take results to 1080p. Released checkpoints include CFG-distilled versions and a 480p image-to-video step-distilled model that needs either 8 or 12 sampling steps, alongside sparse attention, cache-based speedups, and FP8 matrix kernels. Plan on Linux, Python 3.10 or newer, and an NVIDIA GPU with roughly 14 GB of VRAM when offloading is enabled; Flash, Sage, or flex-block attention are optional extras. Diffusers, ComfyUI, and LightX2V support it, and training plus LoRA scripts arrived in December 2025. Weights fall under the Tencent Hunyuan Community License, which excludes the EU, UK, and South Korea and requires a separate deal for services above 100 million monthly users.