Can we make max_size=512 = set-able by user or 512 mandatory? Also can we save every generated image into outputs folder

[FurkanGozukara]

In the Gradio APP atm it is hard-coded to resize into 512

Can you let user to change it in the Gradio interface if possible?

Or 512 is mandatory? thank you

I also don't see generated images are not saved anywhere. Can you auto save every generated image into outputs folder?

[zliucz]

Hi, there. Thanks for your interests on our project :). I've watched many of your videos, and many of them are inspiring.

Auto-resizing to 512 is not mandatory. You may check issues#36. Your suggestion to let user change the resizing resolution is brilliant. We'll adjust it in our next update.

The generated images will not be saved automatically. You may click the download icon to manually download it.

[FurkanGozukara]

@zliucz can you add auto save? i certainly can fork and add but i really prefer to keep using your repo :)

and thank you so much. both features are asked by many people to me

i plan to make a tutorial for MagicQuill

[zliucz]

Of course! Thanks for your suggestion:).

We’ll update these functions in our next version.

[zliucz]

Hi, there @FurkanGozukara. We have updated our code to implement these two features you've mentioned. Thank you again for your brilliant suggestions. Pull our latest code to try it out :).

[FurkanGozukara]

@zliucz awesome ty

bigger res uses more vram though :D

do you think any quant or optimization possible?

[zliucz]

@FurkanGozukara Indeed. We have already applied 4bit-quantization to the Draw&Guess LLaVA. We'll look into possible optimization strategies :).

[FurkanGozukara]

@zliucz awesome

by the way i made it work on runpod and massed compute via cloudflared will make a tutorial soon hopefully

[zliucz]

I'm glad to hear that, @FurkanGozukara. Let us know if you have further suggestions for our project.

I have always been a big fan of your YouTube channel ;) And I started my journey in generative AI by watching your tutorial about fine-tuning a LoRA. Thanks for your contribution to the community.