Info about SD development. (Patreon)

Hi all, following the results of the pool, I did not expend time (for now) on improving the seamless code.

I'm also giving a shot on using Google Forms to collect suggestions and bug reports:
https://forms.gle/Hq8e8jJNMo2CD3mZ6
This may help me to organize all suggestions and bugs that need fixing.

I'm currently very focused on making Lora work. Lora is more complex to implement that I originally though. More info on that at the end of the post.

I'm still doing some tests, but its very possible that I will drop hypernetworks option, Lora is pretty much hypernetworks on steroids. It work better, is more flexible and give better results. Plus hypernetworks are getting dropped by the libraries, so it's getting harder to implement without breaking some new speed optimizations.

I'm still gonna try a little more to make Lora work for this next build, but if progress is slow, I may focus more on bug fixing and delay Lora for the next build.

Other than that been fixing a lot of bugs reported for you all, thanks for all the feedback, this really help me find all the problems.

Will try to add some improvements. Like the possibility to comment a line, sometimes you want to do some tests with the prompt, so my idea is just to copy the prompt to a new line and comment the original one, so you don't lose information.

I'm also working on a small update to Whisper, to improve timing on the subtitles.

After I manage to complete most of the SD update, what to give a shot on a GUI for one of those GPT open source codes.

About Lora:
Lora (for the best of my currently knowledge) can replace almost all layers from Unet and Vae, it also add a new variable for itself if the replacement occur.

Nothing incredible complicated. The problem is that this is really new, so there is no standard procedure on how to do it, so everyone have it's solution for it, it's own code on how to load this file to that model format.

The version that is getting popular is the .safetensor one that start the dict with lora_unet and lora_vae, this .safetensor require that the model use the .ckdr version of Unet and .pt version of Vae.

There is no mistery on how the code work, it just take work for now to make it all work for my currently implementation of Unet and Vae. I'm sure that soon there will be an official integration for Lora using my version of Unet and Vae, but right now I'm doing a lot of this integration myself.

Message from nick barry:

Greetings folks - I just finished this animated short film (2mins) using SD images generated via GRisk UI. Using 1.4 and 1.5. Check it out. Photoshop, After Effects, Premiere utilized. 2nd in the series "This Art Must Exist" I am doing upload quality testing. It is unlisted and will go public this weekend along with #1 in the series. Creating this art was not possible without the GRisk UI. Thanks- https://youtu.be/WsxBJ9LoQVM