Summary

HunyuanVideo-1.5 is a lightweight, high-performance video generation model offering state-of-the-art quality with an accessible 8.3B parameters, designed for consumer GPUs. It supports both text-to-video (T2V) and image-to-video (I2V) generation, lowering barriers for developers and creators.

How It Works

The model features an 8.3B-parameter Diffusion Transformer (DiT) with a 3D causal VAE. Its core innovation, Selective and Sliding Tile Attention (SSTA), prunes computations to accelerate inference. It incorporates meticulous data curation, glyph-aware text encoding, and a multi-stage progressive training strategy for enhanced motion coherence and visual quality.

Quick Start & Requirements

• Installation: Clone repo, pip install -r requirements.txt, pip install tencentcloud-sdk-python. Flash Attention, Flex-Block-Attention, SageAttention recommended for performance.
• Prerequisites: Linux, Python 3.10+, NVIDIA GPU (14GB+ VRAM recommended with offloading).
• Running: Use torchrun --nproc_per_node=<N> generate.py .... Pretrained models require separate download.
• Links: GitHub: https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5.

Highlighted Details

• Architecture: 8.3B DiT with 3D VAE, achieving significant compression.
• SSTA: Selective and Sliding Tile Attention provides $1.87 \times$ speedup for 10s 720p synthesis vs. FlashAttention-3.
• Enhancements: Includes an efficient few-step super-resolution network upscaling to 1080p.
• Capabilities: Demonstrates strong instruction following, cinematic aesthetics, text rendering, and physics compliance via advanced prompt rewriting.
• Performance: Optimized inference with CFG distillation and sparse attention, tested on 8 H800 GPUs.

Maintenance & Community

Community contributions are encouraged (e.g., ComfyUI plugins). WeChat and Discord channels are available. Acknowledges open-source contributions from Transformers, Diffusers, HuggingFace, and Qwen-VL.

Licensing & Compatibility

The license type is not specified in the provided README text.

Limitations & Caveats

Distillation and sparse attention models are noted as "coming soon." Diffusers support is not yet implemented. Primary environment appears to be Linux.