hydit-live

A simple gradio interface for local image generation with HunYuan-DiT.

Background Information:

https://arxiv.org/abs/2405.08748 (Original HyD publication)

https://huggingface.co/Tencent-Hunyuan/HunyuanDiT (Main HuggingFace Repo)

In a nutshell, HyD offers impressive capabilities due to its bilingual training dataset and ability to take long, detailed natural language descriptions as prompts.

Installation and use (Tested with Win11/Python 3.10.9/CUDA 12.2):

It is recommended to create a virtual environment first! In powershell:

python -m venv venv

Then, activate the enviroment

.\venv\Scripts\activate.ps1

and run

pip install -r requirements.txt
app_opt.py

to launch the gradio session in your browser. The session creates a share link by default for use over the internet.

Recommended settings:

Inference Steps: 30-50

Guidance Scale: 5-7 (use integers only!)

Obligatory cat sample @ 50 Steps, Guidance = 7, Random Seed:

Prompt: Draw a multidimensional android feline high admiral onboard a latent spacecraft as character portrait in the style of Tetsuya Nomura from circa 1997. Make the portrait look like an official promotional poster for a video game.

You can prompt more like chatting with an LLM-- HyD does not require SD/SDXL style comma separated prompts (feline high admiral onboard a latent spacecraft, character portrait in the style of Tetsuya Nomura, 1997, official promotional poster), but you can do so if you wish.

Notes:

Tested on NVIDIA RTX 40XX, Windows 11.

Coming Soon:

-TensorRT version (large speedup on NVIDIA RTX)

-Apple Silicon version (in alpha)