underground club

hunyuan video:

Thanks to the open-source 'hunyuan' model, I had the opportunity to train this LORA. This is my first time training a video LORA, and although there are still many things I don't fully understand, the final result turned out better than I expected. You can refer to my previous FLUX and XL prompts, which mainly include themes, scenes, lighting, and backgrounds.

The example above is not an image; it's a simple video that needs to be clicked to play. It's made up of several short videos. The semantic understanding is quite good, and I hardly need to keep filtering, as the generation is quite easy.

I used a 512x512 resolution. You can increase the quality for better results, but the speed will be relatively slower.

If you haven't used 'hunyuan' before, here's the official guide and model download link.：https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/，Then simply add a 'LORA' node to the official workflow as usual.

FLUX 1 D :

Flux has better language understanding, better picture quality details, and everyone can try different gameplay

SDXL:

Improved semantic understanding, able to generate more interesting content,

Resolution sampling, etc. is recommended to be used according to the check XL used,

Sometimes if the content is not ideal, besides changing the prompt words, It's best to try different check models,

I often use Dreamshaper XL/Jumpernaut XL/LEOSAM HelloWorld XL for large models,

I haven't tested any other check, everyone can try more,

Lora weight recommendation 0.5-0.7,

The training dataset is mainly based on the scene, and there are not many people taking photos. Therefore, if you want to take photos of people, the weight should be reduced to below 0.5.

SD1.5 model:

Generate a LORA model of the underground club,

Model recommended realistic class check model,

Size recommended 768×512,

High resolution repair is recommended,

Can work with some LORAs for interesting effects