New HunyuanVideoGP v2 even Faster generations and support for low RAM / low VRAM computers

[deepbeepmeep]

I have just released v2 of HunyuanVideoGP.

Thanks to a complete refactoring which required to rewrite entirely the widely used safetensors library, you can now generate videos on low end systems (I have managed to get it to run on a 16 GB RAM Linux running inside WSL with an old RTX2080Ti 11 GB of VRAM).

On high end systems (RTX 3090/4090) it is even faster than before thanks to optimisations that you won’t find anywhere else (including ComfyUI) such as asynchronous transfers and leveraging reserved RAM.

I have also added support for Py torch compilation to get speeds up of 50% on Linux systems.

Last but not least, it is easy to use :

• Full Gradio app, no need for ComfyUI
• Fully functional on Windows (although you won’t be able to use Sage attention or PyTorch compilation without WSL)
• Models that required are automatically downloaded. You can even try out the Fast HunyuanVideo model
• Support for quantized models. Either on the fly quantization or prequantized model.

You can find it here :

https://github.com/deepbeepmeep/HunyuanVideoGP

I will be happy to get your feedbacks.

[FurkanGozukara]

@deepbeepmeep so awesome

are you planning to add lora support? kohya has lora trainer also onetrainer has too

please also add number of generations (so people can set to generate 10 videos with random seed each time) , multi-line prompting (so when enabled each line is a new prompt), open in browser by default or via args, enable gradio live share via args , remove server_name and server_port and make them only if passed via args otherwise it causes problem when 7860 is in use,

[deepbeepmeep]

I guess support for pretrained Lora should be quite simple. Is this your request ? Do you have any example to provide ? Training Lora would require much more work.
The rest of your requests should be easy to implement.

[FurkanGozukara]

@deepbeepmeep thanks

not training but just using trained loras. kohya and onetrainer already making trainer

[deepbeepmeep]

It would save me a lot of time for testing if you could provide one or two pretrained loras known to work with HunyuanVideo.
Thanks

[FurkanGozukara]

It would save me a lot of time for testing if you could provide one or two pretrained loras known to work with HunyuanVideo. Thanks

sure here https://civitai.com/models/1034630/hunyuan-video-arnold-schwarzenegger-lora

[cchance27]

civit has a bunch of them for testing, theyre getting more popular

[deepbeepmeep]

thank you

[deepbeepmeep]

So I have just released version 3.0 with Lora support and mutltiple prompts / multiple generations and extra command line options as requested by FurkanGozukara. I have also improved the user interface so that you can abort a generation and preview already generated videos. Everything is still for the memory poor : only 32 GB of RAM for Lora (versus 48-64 for kohya or onetrainer) and blazingly fast if you turn on all the optimisations (pytorch compilation, sage attention, Tea Cache and profile between 1 and 4).
I would be happy to get your feedback.

[FurkanGozukara]

@deepbeepmeep awesome

i plan to research lora training + how to use

i see that you have --lora-weight path1 path2 ... : list of Loras Path

i think it should read all loras put into loras folder and you can select it from gradio interface

that would be best usage case

i had made that in instantid gradio i developed

[deepbeepmeep]

Hi. I will look into that however it may not be as simple as it requires to unload Lora without restarting the server.
If I am not mistaken @FurkanGozukara you are a VIP/moderator on reddit, I would appreciate if you could give a word about HunyuanHGP. it would enlarge its audience and help me improve it with more feedback. Many thanks in advance.

[FurkanGozukara]

Hi. I will look into that however it may not be as simple as it requires to unload Lora without restarting the server. If I am not mistaken @FurkanGozukara you are a VIP/moderator on reddit, I would appreciate if you could give a word about HunyuanHGP. it would enlarge its audience and help me improve it with more feedback. Many thanks in advance.

Sure i plan to cover with a tutorial

at worst case you can delete the model from memory and reload with new LoRA but there could be already load and unload methods - it exists in other pipelines usually

[deepbeepmeep]

@FurkanGozukara I will try to implement your Lora requirements week next as I have been busy releasing Cosmos1GP.

[FurkanGozukara]

@deepbeepmeep awesome ty

[FurkanGozukara]

@deepbeepmeep also great work with Cosmos1GP i started following that repo as well