DAINAPP

Stable Diffusion got a Github

If you been following the AI projects for Image Generation, you must had at least head of Stable Diffusion. It let create images like this homer thanos from a test of mine:

Well, they just got a Github:

2022-08-10 17:19:03 +0000 UTC View Post

Been working like crazy in the last few days.

Mostly focused on improving/fixing Rife-App in every way I can. Between training also learning more about gans and testing my manga coloration model.

Those are my scripts from last week training the new Rife model, each have about 50h of training, with 4 to 5 modifications. Totalling about 900h o...

Rife-App 3.30

Here we go, the next update for Rife-App.

Pretty much it have new models for animation and Life-Action.
I was gonna to build this version on Saturday, but my car broke down and sending it to get fix took mostly of my time, so sorry for the delay.

Live Action model is ready.

So both the anime and Live action models are ready, is hard to know see the difference via examples, but here are two small examples of the new model
Example 2:
https://streamable.com/jpbuvp

Now I just need to build the new version and upload it.

Soon: Update on Rife-App

Hi all, been adding stuff on Rife-App that I learned while training the coloring model.

Right now the new model is already a lot faster than the current model and give better results. 

Currently the anime model is almost done training and then will do the Live action model, then the release it here =)

Anime Shading and coloring Pre-Alpha 0.1

Important Stuff:

• The color model is taking a lot of memory exponentially with the image resolution. I'm already fixing this on the next version of the model. If you can get Cuda of memory you an try to use it on the CPU, but it might freeze the computer if the resolution is to big.
• If you select both options to shade and color the image it will generate an ...
  2022-06-10 00:33:46 +0000 UTC View Post

Anime shading + coloring later tonight

After 200h or 300h of testing and training, finally the coloring model is doing something.

It was really easy to make a model that put a nice red shader on top of all images, but a model that select a few colors and throw around it was really hard.

There is a lot of improvement already on my mind to test and improve the shader and coloring model. 

Some of the stuff I'm al...

Sorry for the delay on the coloring build.

I would like to finish the build with a decent coloring model. So I been non-stop trying to make the model better, improving this and that each time. The model right now kinda of give some colors to the images, but I would like it to really be good first. 

Here are all my tests so far, each have at least 4h of training, some got to 20h.

If I can't manage to improve it in the ne...

Oh this is cool.

Pixelart generator.

Time to do a new build (once the anime coloring one is complete)

Still didn't look at the libraries it use, there is a few out there that are a pain to add to a .exe build.

But you can access the Notebook here:

2022-06-02 19:51:32 +0000 UTC View Post

Stuff

• The interface for the anime shading/coloring is already working with the shading model, just need to do the coloring model and build it.
• Every day I been improving the coloring model, it starting to show some decent results.
• Once the coloring model is good enough, will train a model to color BW videos.
• Started to train a new model for interpolations with the stu...
  2022-06-01 19:45:01 +0000 UTC View Post

Pretty much the first version of the coloring AI

This will be the version that will be on the coloring build that is mostly ready.

There is a ton of improvements that I will add on the next model, but like always it take a long time to test and improve on it, so even if this version have a lot of limitations, it will take a while until the next one is ready.

Finally the coloring AI is starting to work

It's not great yet, but at least is not using a single color to paint all the image.

The AI will always have a favorite color, in this examples is blue, but I'm trying to give it a default warmer color. After that make another AI that let you select a pallet to paint it.

Will also train a model more focused on HQ after this start working better.

Working on the interface for the anime shading build.

Mostly of it was copied from Rife-App lol

Examples on the Sketch->Shading

Here is some results from my AI to add shading to sketch. I'm still working on the color AI, but it's already generating some basic colors as well.
This will also be the base for my Grayscale->Color Video AI.

For now it's more or less a fun toy, if someone care about it, I can make a build to test it, just let me know.

Rife-App with AMD(?) Support.

So here is my first attemp with the AMD card support. If anyone have a AMD card please send me some feedback after testing it.

You just need to select "AMD" on the device list.

The models are also slighter better than the last release.

Amd support on next update, perhaps?

Been doing some tests, Rife-App may have AMD support in the next update.

There still tests to be done, like. It will work on the final build? We will see.

It should be supported on any Windows with DirectX12, but again, it need more tests.

My limited test for now:

Test video time:

• AMD: 47 sec.
• CUDA: 34 sec.

Rife-App 3.10

Pretty much the stuff from the last posts. Some changes in the models, lighter and a little faster.

A new option "Clean Jpeg Artifacts", but is still in alpha, there is a lot to improve on it yet.

Rife-App 3.1 today/tomorrow

Edit: The left area is the original 24fps footage.

All the code is ready, just need some extra changes for the  build.

Model is now lighter and a little faster.

Beta model to clean artifacts, with a few details:

• The artifact cleaner model is kinda of slow for now, but it should not generate the "Out of memory error" in any resolution.
2022-04-11 18:14:54 +0000 UTC View Post

Sharing a link for the fans of Art Generation.

OpenAi just made a page for Dall.e 2, to generate art. Is not released yet, but you already can try to get to the waitlist once is released.

OpenAI stuff all run online, so it should be easy to use until they start charging for it.

Boris was trying some users inputs on twitter, you can check it out:

2022-04-07 11:40:40 +0000 UTC View Post

Example of the Artifact Cleaner

Here an example using the artifact cleaner. For now I trained it on gif artifacts, but will make in the future another one for mp4 artifacts.

Patreon scale up the video, so it may be better to see it using the link:

2022-04-01 14:48:36 +0000 UTC View Post

Rife-App 3.1 soon

I been crazy busy trying to improve some stuff on Rife-App

• Remade some stuff on the model, to use less memory. It's more or less can run 1024X1024 on 4vram now.
• The artifact cleaner is giving some pretty good results, it will be released as well. Since this model need to run on full resolution, I already have working a split script to clean the  frame little by little,...
  2022-04-01 14:20:33 +0000 UTC View Post

More work on Rife-App

Since 3.0 I'm working to soon release 3.1

• Beta for the Artifacts cleanup
• Cleanup on the code, should speed up a little, improving some of the code.
• Better memory for model. I'm trying to make it work with 1024X1024 for 4Vram, is more or less already working.
• Improvement at the edges, that kinda of pull the colors, already made some progress on this.
<...
2022-03-22 20:12:32 +0000 UTC View Post

Working on a artifact removal to add to Rife-App

Result are still WIP as the model is in training.

Before:

https://streamable.com/c20g5b

After:

https://streamable.com/i1r6xa

I also will train a upscaler later on. Since is my code I can finally put all into a single...

Rife-App 3.0 Release

Here we go. Version 3.0 with both Anime and Live action (new name, it's real life model) models.

The idea was to release it after the Anime model was ready (as in the pool result), but I had to re-train it with a few modifications, so it gave me time to finish the Live action model.

This version have pretty much the same options for the last update, plus the new models and bug fixes...

So you guys don't think i'm pulling your leg.

The devil really are in the details. Slowly been managing to improve the model, but each day there is something new to learn, how batch affect the model and what is the best values for the training.

The models that I have trained so far are pretty good, but it's seen that it only that the right value so they can become great. The test last night manage to make the model work a lot better ...

More on Rife-App

Ok, don't kill me, but while finishing the code for the app I kept training the model with some new tests, the test made a great improvement on the model, so I had to train it a little more.

But good news, it should be ready tomorrow. Even more good news, during this time I managed to train the RL model as well, so the next update will have both models.

RL model can still be improve...

Overall Update

Rife-App:

The animation model is done! Horray. Still need to update some code on the software, this new model require some new stuff on the input, but should be ready soon.

Today will start training the RL model while making the build for the anime model.

For this next build, gonna need to remove the RL model, because the code will have lots of chance, no poin...

Rife-App Update

It seen like the new anime model won't stop improving, it must have more than 300h of training and still improving little by little.

So my question for you all. Once the animation model is done, should I release a new build or wait for the RL model to finish training? Since it will take a while to train the RL model after this one.

Some results for Grayscale Images to Color.

I found a dataset of anime girls faces to train the AI, so this is why the results are anime girl faces lol
Those images only have a grayscale image as input, no color hints or anything, and it seen to be working very nicely so far.

My idea is to have a collection of models to do a lot of stuff with art.
I'm also gonna train one for Black and White RL movies. Not sure how well will ...

Coloring starting to work

Feeding to the AI a grayscale image and a few colored dots it starting to show some results:

Plus Original colors:

Equal:

2022-02-22 18:14:26 +0000 UTC View Post